A sea surface methane concentration inversion prediction method and system based on a time series neural network

Abstract

The application relates to a sea surface methane concentration inversion prediction method and system based on a time sequence neural network, and relates to the technical field of marine environment monitoring and processing. The method comprises the following steps: taking acquired water color remote sensing data and environmental parameters as benchmarks, inversely analyzing optical parameters and diffusion attenuation coefficients, further inversely analyzing key driving quantities reflecting the main physical process of methane change on the sea surface, extracting time sequence characteristics with physical meanings, inputting the time sequence neural network model group to invert methane concentration values, and predicting methane concentration values in a certain time period. It can be seen that the application starts from the process of sea surface methane production and sea surface methane concentration change, deeply studies the influence of factors such as optical characteristics, environmental parameters and various driving quantities on sea surface methane production and emission, on the one hand, realizes accurate inversion of the sea surface methane concentration value at a known target time, and on the other hand, can predict the methane concentration value in a future time period, thereby improving monitoring efficiency and applicability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marine environment monitoring and processing, and particularly relates to a sea surface methane concentration inversion and prediction method and system based on a time series neural network. BACKGROUND

[0002] In the related art, atmospheric methane mainly reflects the overall situation of global carbon cycle; sea surface (referred to as sea surface) water body methane directly reveals the biogeochemical processes inside the water body, including light production, decomposition and the key mechanism of methane emission to the atmosphere. Therefore, accurate monitoring of the sea surface water body methane concentration has independent scientific research significance, and is also a necessary parameter for calculating the water-gas interface methane flux.

[0003] In the prior art, the monitoring and analysis of the methane (CH4) concentration in the sea surface (referred to as sea surface) water body of the global ocean still focuses on the sea surface methane concentration and the sea-air flux at a single time point, mainly relying on in-situ instruments to achieve measurement. However, this method has significant limitations. Specifically, the prior art can only measure the methane concentration or the sea-air flux in a certain area, the measurement coverage is small, and large-scale monitoring cannot be achieved. Furthermore, simply measuring the methane concentration and the sea-air flux data at a single time cannot accurately reflect the concentration change trend and emission trend of the sea surface methane, and the production and emission of the sea surface methane are affected by environmental factors such as temperature, and the prior art does not deeply study the sea surface methane production process, the methane concentration inversion result and the prediction result of the emission trend are not accurate.

[0004] It can be seen that the prior art cannot dynamically and accurately invert the concentration of the sea surface methane, nor can it predict the methane concentration value in the future time period. SUMMARY

[0005] The present application provides a sea surface methane concentration inversion and prediction method and system based on a time series neural network, which inverts the key driving quantity related to methane concentration prediction, extracts time series (referred to as "time series") features with clear physical meaning, and then inverts the sea surface methane concentration value at the known target time and predicts the sea surface methane concentration value in the future time period, realizing dynamic and accurate inversion of the sea surface water body methane concentration and being able to predict the emission trend of the methane. The present application has significant advantages in terms of technical innovation, monitoring efficiency and applicability, can provide efficient and reliable data support for global marine methane cycle research and environmental protection, and solves the problem that the prior art cannot dynamically and accurately invert the sea surface methane concentration value and predict the sea surface methane concentration value in the future time period.

[0006] In a first aspect, the present application provides a sea surface methane concentration inversion and prediction method based on a time series neural network, comprising:

[0007] Obtaining water color remote sensing data and environmental parameter information of a target area within a preset time period, the water color remote sensing data including sea surface remote sensing reflectance obtained by a satellite;

[0008] Performing water body optical inversion based on the sea surface remote sensing reflectance to obtain optical characteristic parameters of colored dissolved organic matter and a diffusion attenuation coefficient of the water body, the optical characteristic parameters including an absorption coefficient spectrum and a spectral slope at a key wavelength;

[0009] Performing analysis and processing based on the environmental parameter information and obtained methane concentration information to obtain dissipation information and a methane oxidation rate, the dissipation information including a sea-air exchange flux and a bubble flux;

[0010] Analyzing a terrestrial input index according to the spectral slope, and inversely analyzing a production efficiency of sea surface methane in situ, including a photo-production efficiency and a biological production efficiency, according to the optical characteristic parameters, the diffusion attenuation coefficient and the environmental parameter information;

[0011] Taking the production efficiency of sea surface methane in situ, the dissipation information, the terrestrial input index, the methane oxidation rate and the environmental parameter information as inputs, performing feature extraction of a basic time sequence and a special time sequence through a time sequence feature extraction module to obtain time sequence feature information;

[0012] Constructing time sequence data based on the production efficiency of sea surface methane in situ, the dissipation information, the terrestrial input index and the methane oxidation rate, combining the time sequence feature information, and analyzing short-term and long-term features through a preset time sequence neural network model group to obtain a first sea surface methane concentration value and a second sea surface methane concentration value predicted within a specified time;

[0013] The time sequence neural network model group includes a first model and a second model, the first model inversely analyzes the first sea surface methane concentration value based on input data, the second model predicts the second sea surface methane concentration value based on input data, initial acquisition of the methane concentration information is obtained by inverse analysis based on climatological statistics, and is continuously updated during running of the time sequence neural network model group, and initial acquisition of the methane oxidation rate and the dissipation information is obtained by inverse analysis based on initial methane concentration information, and is continuously updated during running of the time sequence neural network model group.

[0014] Optionally, performing water body optical inversion based on the sea surface remote sensing reflectance to obtain optical characteristic parameters of colored dissolved organic matter and a diffusion attenuation coefficient of the water body includes:

[0015] Performing principal component analysis and color cluster classification according to the marine remote sensing reflectance, and performing multiple linear regression calculation to obtain an absorption coefficient spectrum and a spectral slope of colored dissolved organic matter;

[0016] According to the marine remote sensing reflectivity, principal component analysis and cluster analysis are performed, and an optimization algorithm is used for optimization processing, so as to obtain the diffuse attenuation coefficient of the water body.

[0017] Optionally, according to the marine remote sensing reflectivity, principal component analysis and color cluster classification are performed, and multivariate linear regression calculation is performed, so as to obtain the absorption coefficient spectrum and spectral slope of the colored soluble organic matter, including:

[0018] According to The sea surface remote sensing reflectivity is logarithmically transformed, and according to The logarithmically transformed sea surface remote sensing reflectivity is standardized to obtain a first remote sensing reflectivity;

[0019] The first remote sensing reflectivity is input into a pre-trained feature vector matrix, and according to [PC1 i , PC2 i , PC3 i ] = R' rs (λ) i ×U, principal component analysis is performed to calculate three principal component values;

[0020] Based on the three principal component values, classification and determination of the marine color cluster are performed to determine the target cluster number to which each principal component value corresponds in the marine color cluster;

[0021] Taking the target cluster number as a reference, a multivariate regression linear equation is called, according to ln(a g (λ)) = β0(λ) + β1(λ) * PC1 i + β2(λ) * PC2 i + β3(λ) * PC3 i , the absorption coefficient corresponding to the sea surface remote sensing reflectivity at each wavelength is calculated, and the complete absorption coefficient spectrum a g (λ) is calculated;

[0022] Taking the first remote sensing reflectivity at each wavelength as input, according to ln(S 275-295 ) = α + β * ln[R rs (443)] + γ * ln[R rs (488)] + δ * ln[R rs 531)] + ε * ln[R rs (555)] + ζ * ln[R rs (667)], the spectral slope S 275-295 of interest is analyzed and calculated;

[0023] wherein, R' rs (λ) iwherein, R is the first remote sensing reflectance, U is a feature vector matrix obtained through algorithm training, PC1, PC2 and PC3 are principal component values, α, β, γ, δ, ε and ζ are regression coefficients, mean and std are parameters predetermined in the algorithm.

[0024] Optionally, principal component analysis and cluster analysis are performed on the marine remote sensing reflectance, and an optimization algorithm is used for optimization processing to obtain the diffuse attenuation coefficient of the water body, including:

[0025] The satellite band corresponding to the marine remote sensing reflectance is analyzed, and standardization processing is performed based on the satellite band to obtain a second remote sensing reflectance belonging to a preset wavelength;

[0026] Water body type judgment is performed according to the second remote sensing reflectance, and an optimization algorithm is selected according to the water body type judgment result;

[0027] According to The second remote sensing reflectance is normalized to obtain normalized data X(λ);

[0028] The normalized data X(λ) of each wavelength is combined with the feature vector of the optimization algorithm as input, and four principal component values are analyzed and calculated according to PC j =e j1 X(412)+e j2 X(443)+e j3 X(490)+e j4 X(510)+e j5 X(555)+e j6 X(670).

[0029] The four principal component values are input into a multiple linear regression equation model of the optimization algorithm, and the wavelength diffuse attenuation coefficient of the wavelength corresponding to the four principal component values is calculated according to ln[Kd(λ)]=α+βPC1+γPC2+δPC3+∈PC4.

[0030] According to The wavelength diffuse attenuation coefficients are exponentiated to obtain the real diffuse attenuation coefficient Kd(λ).

[0031] wherein, λ is the data wavelength, σ R (λ) is a standard deviation predetermined according to the water body type, is a mean value predetermined according to the water body type, e j is a given feature vector.

[0032] Optionally, the environmental parameter information and the obtained methane concentration information are analyzed and processed to obtain dissipation information and a methane oxidation rate, including:

[0033] Taking the temperature, salinity and water depth information in the environmental parameter information and the methane concentration information as inputs, a preset random forest regression model is used to inverse the methane oxidation rate;

[0034] Taking the wind speed information in the environmental parameter information as input, according to Calculate the gas transfer rate k;

[0035] Taking the gas transfer rate k as input, according to F diff = k ([CH4] water - [CH4] air ), the sea-air exchange flux F diff of methane is calculated.

[0036] According to Calculate the bubble flux;

[0037] Wherein, is the wind speed above the sea surface, Sc is the Schmidt number, [CH4] water is the methane concentration in the sea surface water body, [CH4] air is the methane concentration above the sea surface.

[0038] Optionally, according to the spectral slope, the terrigenous input index is analyzed, including:

[0039] According to The characteristic information of the spectral slope S is analyzed;

[0040] Taking the characteristic information as input, according to f 陆源 =

[0041] [ln(S 样品 )-ln(S 陆源端元 )] / [ln(S 海洋端元 )-ln(S 陆源端元 )] The terrigenous input index f 陆源 is quantitatively calculated.

[0042] Wherein, the characteristic information includes the characteristics of the flat falling absorption curve and the steep falling absorption curve, the absorption curve characteristics of the sample with the most flat falling absorption curve and the smallest spectral slope in the sample space, which represents S 陆源端元 , the absorption curve characteristics of the sample with the most steep falling absorption curve and the largest spectral slope in the sample space, which represents S 海洋端元 .

[0043] Optionally, according to the optical characteristic parameter, the diffuse attenuation coefficient and the environmental parameter information, the in-situ production efficiency of the sea surface methane is inversely analyzed, including:

[0044] According to the environmental parameter information and the absorption coefficient spectrum, a spectral quantum yield spectrum related to methane photo-production is inversely analyzed, and an underwater spectral scalar irradiance is obtained by a radiation transfer model;

[0045] Based on the diffuse attenuation coefficient and the underwater spectral scalar irradiance, exponential attenuation of the underwater spectral scalar irradiance at a specified depth is analyzed, and a photon absorption rate at the specified depth is analyzed in combination with the absorption coefficient spectrum and the apparent quantum yield;

[0046] Based on the photon absorption rate, spectral range integration and vertical integration are performed to obtain a methane photo-production efficiency;

[0047] According to the optical property parameters and the environmental parameter information, a biological production efficiency of methane is inversely analyzed.

[0048] Optionally, the in-situ production efficiency, the dissipation information, the terrestrial input index, the methane oxidation rate, and the environmental parameter information are input into a time series feature extraction module to perform basic time series and special time series feature extraction, and time series feature information is obtained, including:

[0049] The in-situ production efficiency, the dissipation information, the methane oxidation rate, and the terrestrial input index are subjected to basic time series feature analysis and special time series feature analysis by the time series feature extraction module, and first time series feature information is obtained;

[0050] The environmental parameter information is subjected to special time series feature analysis by the time series feature extraction module, and second time series feature information is obtained, and time series feature information is generated based on the first time series feature information and the second time series feature information;

[0051] The first time series feature information includes short-term dynamic characteristics, average state characteristics, recent state characteristics, and long-term trend item characteristics of sea surface methane concentration, and the second time series feature information includes related characteristics of sea surface methane concentration variation affected by wind speed.

[0052] Optionally, before analyzing short-term and long-term characteristics by the preset time series neural network model group, the following steps are further included:

[0053] Initial samples are obtained from a preset database source, and the initial samples include sea surface remote sensing reflectance samples, environmental parameter samples, optical parameter samples, diffuse attenuation coefficient samples, downwelling irradiance samples, and photosynthetically active radiation samples;

[0054] The initial samples are subjected to standardization processing and time series data construction processing to obtain time series data samples with time series continuity, and the time series data samples are divided into a training set and a validation set;

[0055] The time sequence feature extraction module and the physical quantity derivation module are constructed.

[0056] The physical quantity derivation module is trained based on the training set, a key driving quantity related to the sea surface methane concentration change is derived, a time sequence sample is constructed, and the time sequence feature extraction module is trained for baseline time sequence feature and special time sequence feature extraction with the key driving quantity as input to obtain a time sequence feature sample.

[0057] An initial first model and an initial second model are constructed, and the time sequence feature sample and the time sequence sample are inputted, the initial first model and the initial second model are both time sequence neural network models, and form an initial time sequence neural network model group.

[0058] The initial time sequence neural network model group learns the change characteristics associated with the actual physical process from the input time sequence feature sample and the input time sequence sample, learns the short-term features and the long-term features from the input time sequence data, automatically extracts and encodes the key information in the input time sequence through parameter adjustment and iterative optimization, gradually constructs the feature representation of the time sequence for the inversion and prediction of the sea surface methane concentration, and outputs the inversion result and the prediction result.

[0059] The inversion result and the prediction result are respectively verified and evaluated through the verification set, the parameters of the initial time sequence neural network model group are optimized until a trained time sequence neural network model group is obtained.

[0060] The initial first model outputs the inversion result, the initial second model outputs the prediction result, the features extracted in each model in the initial time sequence neural network model group are black-box learning, which is used for effectively extracting feature parameters to improve the prediction accuracy, the time sequence feature sample includes the extracted time sequence features with clear physical meaning, which is used for enhancing the model prediction performance and providing scientific interpretability.

[0061] In a second aspect, the present application provides a sea surface methane concentration inversion and prediction system based on a time sequence neural network, comprising:

[0062] A data acquisition module is configured to acquire water color remote sensing data and environmental parameter information of a target area within a preset time period, wherein the water color remote sensing data includes sea surface remote sensing reflectivity acquired by a satellite.

[0063] An optical inversion module is configured to perform water body optical inversion based on the sea surface remote sensing reflectivity to obtain optical property parameters of colored soluble organic matter and a diffuse attenuation coefficient of the water body, wherein the optical property parameters include an absorption coefficient spectrum and a spectral slope at a key wavelength.

[0064] a dissipation and oxidation rate analysis module configured to analyze and process the environmental parameter information and the obtained methane concentration information to obtain dissipation information and a methane oxidation rate, the dissipation information including a sea-air exchange flux and a bubble flux;

[0065] an in-situ production efficiency analysis module configured to analyze a terrestrial input index according to the spectral slope and to inversely analyze an in-situ production efficiency of the sea surface methane according to the optical characteristic parameters, the diffuse attenuation coefficient and the environmental parameter information, the in-situ production efficiency including a photo-production efficiency and a biological production efficiency;

[0066] a feature extraction module configured to take the in-situ production efficiency, the dissipation information, the terrestrial input index, the methane oxidation rate and the environmental parameter information as inputs, to perform basic time series and special time series feature extraction through a time series feature extraction module, and to obtain time series feature information;

[0067] a sea surface methane concentration inversion and prediction module configured to construct time series data based on the in-situ production efficiency, the dissipation information, the terrestrial input index and the methane oxidation rate, to combine the time series feature information, to analyze short-term and long-term features through a preset time series neural network model group, and to obtain a first sea surface methane concentration value and a second sea surface methane concentration value predicted within a specified time;

[0068] wherein the time series neural network model group includes a first model and a second model, the first model is configured to inversely analyze the first sea surface methane concentration value based on input data, the second model is configured to predict the second sea surface methane concentration value based on input data, initial acquisition of the methane concentration information is obtained by inverse analysis based on climatological statistics, and is continuously updated during running of the time series neural network model group, and initial acquisition of the methane oxidation rate and the dissipation information is obtained by inverse analysis based on initial methane concentration information, and is continuously updated during running of the time series neural network model group.

[0069] In summary, the embodiment of the present application uses the water color remote sensing data with time sequence to analyze the optical characteristic parameters of colored soluble organic matter and the diffusion attenuation coefficient of water body. Then, combined with the obtained environmental parameter information and the methane oxidation rate, the key driving quantities with clear physical meaning highly related to the budget process of methane at the sea surface layer, including dissipation information, land source input index and methane in-situ production efficiency, are analyzed. Taking each key driving quantity as a benchmark, the time sequence characteristics are extracted, and finally the short-term characteristics and long-term characteristics are analyzed and predicted through the time sequence neural network model group to inverse the sea surface methane concentration value at the known target time and predict the sea surface methane concentration value in the future time period. It can be seen that the present application starts from the process of sea surface methane production and sea surface methane concentration change, and deeply studies the influence of factors such as optical characteristics, environmental parameters and driving quantities on the sea surface methane production and emission. On the one hand, the sea surface methane concentration value at the known target time is accurately inverted, and on the other hand, the methane concentration value in the future time period can be predicted, thereby improving the monitoring efficiency and applicability. BRIEF DESCRIPTION OF DRAWINGS

[0070] The accompanying drawings, which are incorporated into and form a part of the specification, illustrate an embodiment consistent with the present application and, together with the description, serve to explain the principles of the application.

[0071] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without creative labor.

[0072] Figure 1 A flowchart of a sea surface methane concentration inversion and prediction method based on a time sequence neural network provided by an embodiment of the present application;

[0073] Figure 2 A step flowchart of a sea surface methane concentration inversion and prediction method based on a time sequence neural network provided by an optional embodiment of the present application;

[0074] Figure 3 A flowchart of a CDOM absorption coefficient spectrum inversion based on remote sensing reflectance provided by an optional example of the present application;

[0075] Figure 4 A flowchart of a diffusion attenuation coefficient inversion based on remote sensing reflectance provided by an optional example of the present application;

[0076] Figure 5 A time sequence neural network structure schematic diagram provided by an optional example of the present application;

[0077] Figure 6 A structural block diagram of a sea surface methane concentration inversion and prediction system based on a time series neural network is provided for the embodiments of the present application. DETAILED DESCRIPTION

[0078] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below in detail with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0079] In the related art, most of the existing technologies only focus on the measurement results of experimental data, without in-depth research and analysis of the production and emission process of sea surface methane. Therefore, it is not only impossible to accurately predict the sea surface methane concentration and its corresponding emission trend, but also difficult to promote and apply in a large range.

[0080] To solve the above technical problems, one of the technical concepts of the present embodiment is to deeply study the internal dynamic pattern of sea surface methane in the "disturbance-accumulation-release" process, analyze the key driving quantities related to the change of methane concentration with clear physical meaning, and extract the time sequence characteristics, and input them into the time series neural network model group for inversion and prediction. On the one hand, the dynamic and accurate inversion of the sea surface water methane concentration value at the known target time is realized, the defects of the existing technology in the inversion of the sea surface water methane concentration are solved, and the monitoring efficiency and applicability are improved. On the other hand, the sea surface methane concentration value in the future time period can be accurately predicted, which can provide efficient and reliable data support and theoretical support for global marine methane circulation research and environmental protection.

[0081] For the understanding of the embodiments of the present application, further explanation and description will be made in combination with the drawings and specific embodiments. The embodiments do not constitute a limitation on the embodiments of the present application.

[0082] Figure 1 A flowchart of a sea surface methane concentration inversion and prediction method based on a time series neural network is provided for the embodiments of the present application. In specific implementation, the present embodiment aims to solve the problems of insufficient monitoring range, insufficient parameter integrity, and missing time sequence in the prior art, and proposes an efficient and accurate sea surface methane concentration inversion and prediction method to support accurate analysis of marine carbon cycle and methane flux. As shown in Figure 1 The sea surface methane concentration inversion and prediction method based on a time series neural network provided by the embodiments of the present application can specifically include the following steps:

[0083] Step 110, obtaining water color remote sensing data and environmental parameter information of a target region in a preset time period.

[0084] The water color remote sensing data comprises sea surface remote sensing reflectance obtained by a satellite.

[0085] In the embodiment, the water color remote sensing data is remote sensing data obtained by a water color satellite. The environmental parameter information includes but is not limited to sea surface water temperature, sea surface wind speed and other key parameters.

[0086] In the related art, the remote sensing technology can be applied to the inversion of atmospheric methane concentration due to its monitoring capability. However, in the prior art, there is no technical solution to invert the sea surface water body methane concentration. The reason is that the atmospheric methane remote sensing inversion is mainly based on the specific absorption of methane gas at specific wavelengths in the infrared band, and the signals of other optically active components in the sea surface water body mask the weak signal of trace methane. The remote sensing inversion method based on the optical and even electromagnetic characteristics of methane is invalid, and only the key mechanisms of methane related marine biogeochemical processes can be used to invert the concentration of sea surface methane through more indirect biogeochemical modeling methods. These related biogeochemical processes include in-situ production, oxidation and methane escape to the atmosphere. Therefore, due to the great difference between the inversion of atmospheric methane concentration and the inversion of sea surface water body methane content, the prior art has not yet been able to accurately invert the sea surface water body methane concentration value.

[0087] The embodiment is to realize large-scale and continuous methane concentration monitoring in a certain area, even on a global scale. Water color remote sensing data of a specified area within a certain period of time is obtained to invert and predict the sea surface methane concentration value.

[0088] In step 120, water body optical inversion is performed based on the sea surface remote sensing reflectance to obtain the optical characteristic parameters of the colored dissolved organic matter and the diffusion attenuation coefficient of the water body.

[0089] The optical characteristic parameters include the absorption coefficient spectrum and the spectral slope at the key wavelength.

[0090] In a specific implementation, methane is a target inversion parameter that is affected by the lag and accumulation effect of some environmental parameters in time, and single time point data cannot accurately invert the actual parameter condition.

[0091] Therefore, in the embodiment, the absorption coefficient spectrum and the spectral slope of the colored dissolved organic matter / organic matter (CDOM) in the sea surface water body are inverted and analyzed using the sea surface remote sensing reflectance within a certain period of time, and the diffusion attenuation coefficient of the water body (also referred to as the downwelling diffusion attenuation coefficient of the sea surface water body, abbreviated as K d ) is inverted and analyzed to obtain the inversion result as one of the key calculation parameters.

[0092] At step 130, based on the environmental parameter information and the obtained methane concentration information, analysis processing is performed to obtain dissipation information and a methane oxidation rate.

[0093] The dissipation information includes a sea-air exchange flux and a bubble flux, the initial acquisition of the methane concentration information is obtained based on climatological statistical value inversion analysis, and is continuously updated during the running process of the time series neural network model group, and the initial acquisition of the methane oxidation rate and the dissipation information is obtained based on initial methane concentration information inversion analysis, and is continuously updated during the running process of the time series neural network model group.

[0094] In a specific implementation, considering that the methane concentration inversion is affected by the hysteresis and accumulation effect of environmental factors such as wind speed and water temperature, the embodiment analyzes the related parameters that affect the analysis of the methane concentration content, and when inverting the methane concentration, not only uses remote sensing data and environmental parameters as input, but also further calculates related key driving quantities with clear meanings based on these parameter variables. The key driving quantities include: methane in-situ production efficiency, sea-air exchange flux, bubble flux, methane oxidation rate, and land source input index.

[0095] In the embodiment, the methane oxidation rate, the sea-air exchange, and the bubble flux all affect the prediction of the methane concentration value, so in the embodiment, the sea surface water temperature, salinity, and wind speed in the environmental parameters are used to analyze the methane oxidation rate, the sea-air exchange flux (CH4 flux via air-sea exchange), and the bubble flux (bubble-mediated CH4 flux) in combination with the methane concentration.

[0096] In a specific implementation, the embodiment can use the Wanninkhof (2014) formula to calculate the sea-air exchange flux, and the input parameters include wind speed (which can be obtained by remote sensing) and methane concentration.

[0097] In a specific implementation, the embodiment can use the random forest regression model (containing 105 trees, with a maximum of 167 branches) established by Mao et al. (Mao et al., 2022) to invert the sea surface methane oxidation rate. The input of the model includes methane concentration, temperature, salinity, and water depth.

[0098] In the initial online stage, in order to enable the model to have initial running capability, climatological statistical values (i.e., preset sea surface methane concentration and atmospheric methane concentration in a specific month and a specific location) are used as initial input. Subsequently, through model iteration, the sea surface methane concentration value calculated by the model is used as input for subsequent calculation of the sea-air exchange flux.

[0099] The bubble flux can be calculated based on the Monahan formula, the core input of which is the white cap coverage (Monahan and Muircheartaigh, 1980), which is indirectly related to the wind speed, combined with the methane concentration (processed in the same way), to obtain the bubble flux.

[0100] When the wind speed is high, the wave breaking will produce a bubble turbulent system, and the bubbles can carry dissolved gas up to break and enhance the contribution of the non-molecular diffusion process. As a low solubility gas, methane has a high escape efficiency in this reaction process. To accurately analyze the sea surface methane concentration value in the target time period, one of the concerns of the embodiment is the influence of bubble flux on the sea surface methane escape.

[0101] Step 140, analyzing the terrestrial input index according to the spectral slope, and analyzing the in-situ production efficiency of sea surface methane according to the optical property parameters, the diffuse attenuation coefficient and the environmental parameter information.

[0102] The in-situ production efficiency includes photo-production efficiency and biological production efficiency.

[0103] In the embodiment, the terrestrial input index is mainly through the spectral slope S 275–295 Estimation, reflecting the intensity of humus input.

[0104] Most of the prior art solutions focus on the photo-production efficiency of methane. However, the in-situ production efficiency of methane mainly involves photo-production efficiency and biological production efficiency. Therefore, the present embodiment studies the biological production efficiency of methane, which is combined with the photo-production efficiency to determine the in-situ production efficiency of methane.

[0105] In specific implementation, the in-situ production efficiency of methane is affected by environmental parameters. Therefore, the present embodiment mainly uses the absorption coefficient spectrum of CDOM and environmental parameters to analyze the photo-production efficiency of methane, and uses the absorption coefficient spectrum of CDOM, the spectral slope and the environmental parameters to analyze the biological production efficiency of methane.

[0106] Therefore, the present embodiment introduces key parameters such as methane photo / biological production efficiency to accurately reflect the biogeochemical process of sea surface methane, provides more comprehensive and accurate parameter input, and improves the inversion accuracy of methane concentration.

[0107] Step 150, inputting the in-situ production efficiency, the dissipation information, the terrestrial input index, the methane oxidation rate and the environmental parameter information into a time sequence feature extraction module to extract basic time sequence and special time sequence features to obtain time sequence feature information.

[0108] In the embodiment, the time sequence feature extraction module is a module preset for extracting time sequence features from key driving quantities.

[0109] In a specific implementation, based on the physical understanding of the marine methane dynamic process, features with clear physical meanings are extracted from original time sequences (i.e., the aforementioned parameters and key driving quantities) to serve as additional input channels, which helps to improve the model's physical interpretability and highlights the creative labor in the modeling stage.

[0110] Specifically, the in-situ production efficiency, the dissipation information, the terrestrial input index, the methane oxidation rate and the environmental parameter information are input into the time sequence feature extraction module to extract basic time sequence features, mainly including lag variables, window average values, exponentially weighted average values and long-term trends.

[0111] Considering that the change of sea surface methane concentration is often affected by the disturbance-recovery mode of high wind speed events, the embodiment further extracts special time sequence features, also known as special time sequence process features, including wind speed mutation events, dynamic weighted time windows and wind speed event interval features. The basic time sequence features and the special time sequence process features are combined to form time sequence feature information with clear physical meanings, which is used for subsequent model inversion and prediction.

[0112] Step 160, constructing time sequence data based on the in-situ production efficiency, the dissipation information, the terrestrial input index and the methane oxidation rate, combining the time sequence feature information, analyzing short-term and long-term features through a preset time sequence neural network model group to obtain a first sea surface methane concentration value and a second sea surface methane concentration value predicted within a specified time.

[0113] The time sequence neural network model group includes a first model and a second model, the first model is based on input data to analyze the first sea surface methane concentration value, and the second model is based on input data to predict the second sea surface methane concentration value.

[0114] In a specific implementation, the extracted multiple time sequence features are input into a time sequence neural network model group in combination with time sequence data (including in-situ production efficiency, dissipation information, terrigenous input indicators, methane oxidation rate, and environmental parameter information, etc.). The first model and the second model included in the time sequence neural network model group are both time sequence neural network models, which can be used in two types of scenarios: one is to use known time points (such as T0) and time series of input variables and time sequence features before T0 to infer the sea surface methane concentration value at T0, that is, to obtain the first sea surface methane concentration value; the other is to use T0 and input variables before T0 to predict the sea surface methane concentration value in the future (such as T0+1 to T0+3 days in the future), that is, the second sea surface methane concentration value, which is used for trend foresight and early warning. The functions of the two types of models are obtained by training the same network structure respectively, and have high consistency and complementarity. The use of the model can be set according to the use requirements.

[0115] In this embodiment, each time sequence neural network model has learned the variation characteristics associated with the actual physical process in the model training stage, and automatically learns time sequence features (such as short-term local features and long-term global features) of different levels through internal construction in multiple training processes. Therefore, the model can analyze short-term dependence (such as 2-3 days) and long-term trend (such as 4-7 days) based on the input data, extract short-term features (such as rapid fluctuations, sudden events) and long-term features (such as cumulative effects, lag response, slow variable evolution). On the one hand, the sea surface methane concentration value at the known target time is inferred, and on the other hand, the methane concentration value in the future time period is predicted.

[0116] In a specific implementation, in addition to containing data of the day, the water color remote sensing and environmental remote sensing data also contain historical data of a certain time period. This embodiment uses the aforementioned obtained multiple parameters and key driving quantities to obtain time series and extract time sequence features, and predicts through a time sequence model to capture the dynamic change process of the methane concentration, including the cumulative and dynamic influence of the methane concentration on wind speed, temperature, light production rate, etc. Then, the data of the day and the historical data (such as the historical data of the previous 1 to 3 days) are fused to capture the short-term and long-term dynamic change process, thereby effectively improving the accuracy and reliability of the inversion and prediction.

[0117] It should be noted that the known contemporaneous T0 data is used to infer the internal state, parameter or process of the system (here, the methane concentration) for inversion; the unknown contemporaneous T0+n data uses historical data (T0-m to T0) to predict T0+n.

[0118] The time sequence neural network model in the embodiment uses black box learning, but in the input parameters in the model construction stage, part of the time characteristics (such as cumulative effect, lag effect) extracted according to the physical mechanism and other additional parameters with clear physical meaning are included, so that the model not only has good prediction performance in engineering, but also has stronger scientific explanation. This is one of the important innovations that distinguishes it from pure black box models. The role of the time sequence neural network in the scheme is to comprehensively process and identify patterns based on the input data, and finally output the prediction result. The improved model avoids the limitations of pure black box models, and takes into account the engineering practicability and physical scientific rationality.

[0119] It can be seen that the embodiments of the present application use the water color remote sensing data with time sequence obtained to perform water body optical inversion, and analyze the optical characteristic parameters of colored soluble organic matter and the diffusion attenuation coefficient of the water body. Then, combined with the obtained environmental parameter information and the methane oxidation rate, the key driving quantities with clear physical meaning highly related to the budget process of methane at the sea surface layer, including dissipation information, land source input indicators and methane in-situ production efficiency, are analyzed, and the time sequence characteristics are extracted based on each key driving quantity. Finally, the short-term characteristics and long-term characteristics are analyzed and predicted by a time sequence neural network model group, and the sea surface methane concentration value at the known target time is inverted, and the sea surface methane concentration value in the future time period is predicted. It can be seen that the present application starts from the process of sea surface methane production and sea surface methane concentration change, and deeply studies the influence of factors such as optical characteristics, environmental parameters and driving quantities on sea surface methane production and emission. On the one hand, the sea surface methane concentration value at the known target time is accurately inverted, and on the other hand, the methane concentration value in the future time period is predicted, improving the monitoring efficiency and applicability.

[0120] Reference Figure 2 Fig. 1 shows a step flow schematic diagram of a sea surface methane concentration inversion and prediction method based on a time sequence neural network provided by an optional embodiment of the present application. The method can specifically include the following steps:

[0121] Step 210, obtaining water color remote sensing data and environmental parameter information of a target region in a preset time period.

[0122] The water color remote sensing data includes sea surface remote sensing reflectivity obtained by a satellite.

[0123] In practical implementation, accurate monitoring of methane concentration in sea surface waters has independent scientific research significance and is also a necessary parameter for calculating methane flux at the water-air interface. Water methane concentration is influenced by biogeochemical processes; its formation, accumulation, and release into the atmosphere are coupled, jointly affecting the methane content at the sea surface. Therefore, this embodiment primarily uses diverse and comprehensive marine environmental data as input when analyzing sea surface methane content, and fully considers multiple driving forces for inversion, effectively modeling the dynamic changes in sea surface methane and accurately retrieving the sea surface methane concentration.

[0124] In this embodiment, the inversion of sea surface methane concentration mainly utilizes remote sensing data obtained by water color remote sensing technology to fully leverage the advantages of water color remote sensing data in achieving large-scale and continuous sea surface methane concentration inversion.

[0125] Step 220: Based on the marine remote sensing reflectance, perform principal component analysis and color cluster classification, and perform multiple linear regression calculation to obtain the absorption coefficient spectrum and spectral slope of colored soluble organic matter.

[0126] In practical implementation, the CDOM absorption coefficient spectrum represents the absorption characteristics of CDOM at different wavelengths; the CDOM spectral slope represents the trend of CDOM absorption spectrum within a specific wavelength range. This embodiment performs inversion analysis on these two key parameters. Specifically, after obtaining the ocean remote sensing reflectance, this embodiment obtains the CDOM absorption coefficient spectrum and spectral slope through inversion analysis. During the inversion analysis, each band of the ocean remote sensing reflectance is first transformed, and the corresponding remote sensing reflectance for each band is logarithmically transformed and standardized to obtain the standardized remote sensing reflectance. Then, principal component analysis is performed on the remote sensing reflectance to calculate the principal component values ​​of interest (usually the values ​​of the first three principal components, PC). For each principal component value, the ocean color cluster to which it belongs is determined (mainly using a clustering model for determination). Finally, regression coefficients are calculated using the corresponding cluster to obtain the CDOM absorption coefficient spectrum and the spectral slope of interest.

[0127] In an optional embodiment, the above-mentioned principal component analysis and color cluster classification based on the marine remote sensing reflectance, followed by multiple linear regression calculation, to obtain the absorption coefficient spectrum and spectral slope of colored soluble organic matter may specifically include: based on The sea surface remote sensing reflectance was logarithmically transformed, and based on... The logarithmically transformed sea surface remote sensing reflectance is standardized to obtain the first remote sensing reflectance; the first remote sensing reflectance is input into a pre-trained feature vector matrix, and according to [PC1]... i PC2 i ,PC3 i ]=R′rs( λ) i ×U performs principal component analysis to calculate three principal component values; based on the three principal component values, it classifies and determines the target cluster number corresponding to each principal component value in the ocean color cluster; using the target cluster number as a benchmark, it calls the multiple regression linear equation, according to ln(ag(λ))=β0(λ)+β1(λ)*PC1 i +β2(λ)*PC2 i +β3(λ)*PC3 i Calculate the absorption coefficient of the sea surface remote sensing reflectance at each wavelength, and calculate the complete absorption coefficient spectrum a. g (λ); using the first remote sensing reflectance of each wavelength as input, according to ln(S 275-295 )=α+β*ln[R rs (443)]+γ*ln[R rs (488)]+δ*ln[R rs (531)]+ε*ln[R rs (555)]+ζ*ln[R rs (667)], Analyze and calculate the spectral slope S of interest. 275-295 ; where R′ rs (λ) i Let be the first remote sensing reflectance, U be the feature vector matrix obtained through algorithm training, PC1, PC2 and PC3 are principal component values, α, β, γ, δ, ε and ζ are regression coefficients, and mean and std are parameters predetermined in the algorithm.

[0128] In the above formula, U is a determined and fixed eigenvector matrix, and R... rs (λ) i β0 represents the sea surface remote sensing reflectance; β1, β2, and β3 are all pre-trained regression coefficients.

[0129] Reference Figure 3 As shown, in practical implementation, the obtained ocean remote sensing reflectance (Rrs) typically includes multiple bands. This embodiment converts the ocean remote sensing reflectance bands into SeaWiFS standard bands, including: 412 nm, 443 nm, 490 nm, 510 nm, 555 nm, and 670 nm – a total of six bands. For ocean remote sensing reflectance from different satellite bands, interpolation is used to convert it into the aforementioned six bands.

[0130] Then, the Rrs data of each wave band of the input is logarithmically transformed and normalized. To achieve the logarithmic transformation and normalization, the embodiment adopts a logarithmic transformation formula to accurately perform logarithmic transformation on the Rrs data, facilitates the normalization processing, and then accurately achieves the normalization processing by using a normalization formula, so that the absorption coefficient spectrum and the spectral slope can be accurately obtained during the inversion analysis.

[0131] Next, the normalized Rrs data is substituted into the PCA feature vector matrix trained in advance, and the first three principal component scores are calculated, that is, the first three principal component values PC1, PC2, and PC3 are obtained.

[0132] Then, the three PC scores obtained above are input into the marine color cluster classifier (clustering model) determined during the training. The classifier can automatically determine which one of the nine marine color clusters the current sample belongs to according to the values of PC1, PC2, and PC3. The cluster number (one of the nine categories) to which the data belongs is calculated.

[0133] Finally, for the absorption coefficient spectrum, the trained regression coefficients corresponding to the cluster determined in the previous step are called to perform calculation by substituting into the multiple linear regression equation. For each wavelength λ in the range of 275-450 nm, the corresponding CDOM absorption coefficient can be obtained, and a complete CDOM absorption coefficient spectrum is obtained.

[0134] For the spectral slope, the spectral slope value S 275-295 .

[0135] In step 230, principal component analysis and clustering analysis are performed on the marine remote sensing reflectance, and an optimization algorithm is used for optimization processing to obtain the diffuse attenuation coefficient of the water body.

[0136] In the embodiment, the diffuse attenuation coefficient is used to describe the vertical attenuation degree of light in the sea surface water body.

[0137] The embodiment mainly uses the SeaUV algorithm to calculate the diffuse attenuation coefficient in combination with principal component analysis and clustering.

[0138] In a specific implementation, when the diffuse attenuation coefficient is analyzed by inversion, the wavelength of the sea surface remote sensing reflectance data is first converted into a standard wave band, and then a reasonable algorithm is selected and applied to the principal component analysis process to analyze the first four principal component values. Before the principal component analysis, the sea surface remote sensing reflectance can also be normalized for pretreatment, thereby facilitating the principal component analysis.

[0139] Then, the four principal component values analyzed can be subjected to clustering analysis, and the marine optical clustering domain to which the first two principal component values belong is analyzed to improve the subsequent prediction accuracy and adapt to regions with complex optical properties.

[0140] Finally, the four principal component values ​​were analyzed, and the diffuse attenuation coefficients for different wavelengths were calculated using the algorithm's multiple linear regression model, ultimately yielding the true diffuse attenuation coefficient Kd.

[0141] In an optional embodiment, the above-mentioned principal component analysis and cluster analysis based on the marine remote sensing reflectance, followed by optimization processing using an optimization algorithm to obtain the diffuse attenuation coefficient of the water body, may specifically include: analyzing the satellite bands corresponding to the marine remote sensing reflectance, and performing standardization processing based on the satellite bands to obtain a second remote sensing reflectance belonging to a preset wavelength; determining the water body type based on the second remote sensing reflectance, and selecting an optimization algorithm based on the water body type determination result; and based on... The second remote sensing reflectance is normalized to obtain normalized data X(λ); the normalized data X(λ) for each wavelength is combined with the feature vector of the optimization algorithm as input, and then processed according to PC... j =e j1 X(412)+e j2 X(443)+e j3 X(490)+e j4 X(510)+e j5 X(555)+e j6 X(670), analyze and calculate the values ​​of the four principal components; input the four principal component values ​​into the multiple linear regression equation model of the optimization algorithm, according to ln[K d [λ]=α+βPC1+γPC2+δPC3+∈PC4, calculate the wavelength diffuse attenuation coefficients corresponding to the wavelengths of the four principal component values; according to By taking the exponent of each wavelength's diffuse attenuation coefficient, the true diffuse attenuation coefficient K is obtained. d (λ); where λ is the data wavelength, σ R (λ) represents the standard deviation predetermined based on the water body type. e is an average value predetermined based on the type of water body. j Given the feature vector.

[0142] For example, refer to Figure 4 As shown, taking Kd(490) as an example, according to NASA's standard algorithm for Kd(490), Kd(490) is first calculated: where, if Kd(490) < 0.32m -1 If the water is clear, the original SeaUV algorithm can be used; if Kd(490) ≥ 0.32m -1 If the water quality is poor, it is classified as a complex nearshore water body, and the optimized SeaUV algorithm can be used. The threshold for Kd(490) is 0.32m. -1is based on the NASA standard Kd(490) product and can distinguish between open ocean and complex near-shore waters.

[0143] Next, the remote sensing reflectance data of the selected algorithm (original or optimized SeaUV) is taken as the natural logarithm to obtain ln(R rs (λ)), and the logarithmic converted data is normalized to obtain X(λ).

[0144] The first four principal component scores are calculated using the normalized data X(λ) combined with the eigenvectors of the SeaUV algorithm. Preferably, after analyzing the principal component values, the marine optical cluster domain (for example, open ocean domain OCD or dark water domain DWD) to which the first two principal component scores belong can also be determined.

[0145] Finally, according to the first four principal component scores, the K d (λ) of different wavelengths is calculated using a multiple linear regression model, and the calculated result is taken as an index to obtain the real K d value.

[0146] Thus, according to the above calculation process, the K d value of each waveband (λ = 320, 340, 380, 412, 443, and 490 nm) is accurately calculated.

[0147] The optical parameters and diffuse attenuation coefficient of CDOM are accurately inverted in this embodiment, which provides accurate input for subsequent inversion of methane in-situ production efficiency and driving quantity.

[0148] Step 240, based on the environmental parameter information and the obtained methane concentration information, analysis and processing are performed to obtain dissipation information and methane oxidation rate.

[0149] The dissipation information includes sea-air exchange flux and bubble flux, the initial acquisition of the methane concentration information is obtained based on climatological statistical value inversion analysis, and is continuously updated during the operation of the time series neural network model group, and the initial acquisition of the methane oxidation rate and the dissipation information is obtained based on the initial methane concentration information inversion analysis, and is continuously updated during the operation of the time series neural network model group.

[0150] Optionally, the present embodiment based on the environmental parameter information and the obtained methane concentration information for analysis and processing to obtain dissipation information and methane oxidation rate can include: using the temperature, salinity and water depth information in the environmental parameters and the methane concentration information as input, using a preset random forest regression model to invert the methane oxidation rate; using the wind speed information in the environmental parameter information as input, and according to computing a gas transfer rate k; computing a sea-air exchange flux F diff = k ([CH4] water - [CH4] air ), of methane according to F diff = k ([CH4] ; wherein, is a wind speed above the sea surface, Sc is a Schmidt number, [CH4] water is a sea surface water body methane concentration, [CH4] air is a sea surface above methane concentration.

[0151] In this embodiment, the sea-air exchange flux is mainly calculated using the Wanninkhof (2014) formula to calculate the gas transfer rate, wherein the Schmidt number Sc can be calculated based on the temperature and salinity of the sea surface water body. Preferably, may be the wind speed at 10 meters (m for short) above the sea surface. The initial can use the climatology data of two concentration data as the initial input of the model, and the subsequent prediction can be updated as the time series model runs.

[0152] This embodiment uses the white cap coverage of the sea surface to represent the bubble flux. The marine white cap is a direct manifestation of the wave breaking of the sea surface, and can effectively reflect the intensity of the bubble turbulent system generated by wave breaking.

[0153] Step 250, according to the spectral slope analysis of the terrestrial input index, and according to the optical property parameters, the diffuse attenuation coefficient and the environmental parameter information, the in-situ production efficiency of the sea surface methane is analyzed.

[0154] Wherein, the in-situ production efficiency includes photo-production efficiency and biological production efficiency.

[0155] Optionally, the application embodiment according to the spectral slope analysis of the terrestrial input index can include: according to analyzing the characteristic information of the spectral slope S; taking the characteristic information as input, according to f 陆源 = [ln (S 样品 ) - ln (S 陆源端元 )] / [ln (S 海洋端元 ) - ln (S 陆源端元 )], the terrestrial input index f 陆源 is quantitatively calculated; wherein, the characteristic information includes the characteristics of the flat falling absorption curve and the steep falling absorption curve, the absorption curve characteristics of the sample with the most flat falling absorption curve and the smallest spectral slope in the sample space, representing S 陆源端元 , the absorption curve characteristics of the sample with the most steep falling absorption curve and the largest spectral slope in the sample space, representing S 海洋端元 .

[0156] In a specific implementation, a spectral slope S of the CDOM absorption coefficient has a physical meaning for distinguishing water masses and characterizing terrigenous input materials (Fichot and Benner, 2012). The spectral slope S is usually obtained by fitting the exponential slope of the CDOM absorption coefficient with respect to wavelength. Wherein, λ0 is a reference wavelength.

[0157] According to the above formula, the characteristics of the spectral slope can be calculated and analyzed. The characteristics of the spectral slope mainly include a flat declining absorption curve characteristic and a steep declining absorption curve characteristic. Specifically, at a low S value (flat declining absorption curve), terrigenous CDOM (containing high molecular weight humus) is represented; at a high S value (steep declining absorption curve), marine autogenic CDOM (low molecular weight material) or photodegradation products are represented. Based on the characteristics of the spectral slope, it can be used as a quantitative indicator of terrigenous input to calculate the terrigenous input.

[0158] Thus, the embodiment realizes the inversion analysis of three key driving quantities, including the sea-air exchange flux, the bubble flux, and the terrigenous input indicator. In the embodiment, the in-situ production efficiency of the sea surface methane as one of the key driving quantities has important significance in the inversion of the sea surface methane concentration. In addition, the embodiment focuses on the research on the production process of the sea surface methane, and finds that environmental parameters (such as the sea surface water temperature) have a great influence on the methane production. Therefore, the embodiment also introduces the environmental parameters to jointly fit and analyze the in-situ production efficiency of the methane.

[0159] Optionally, the inversion analysis of the in-situ production efficiency of the sea surface methane according to the optical characteristic parameter, the diffuse attenuation coefficient, and the environmental parameter information can include the following sub-steps:

[0160] In sub-step 2501, a spectral quantum yield spectrum related to the photo-production of methane is inversely analyzed according to the environmental parameter information and the absorption coefficient spectrum, and underwater spectral scalar irradiance is obtained through a radiation transfer model.

[0161] In a specific implementation, when the photo-production efficiency of the in-situ production efficiency of the methane is inversely analyzed in the embodiment, three core elements are mainly involved, that is, the underwater spectral scalar irradiance E o (λ, z), the marine surface spectral diffuse attenuation coefficient K d (λ), and the apparent quantum yield of the methane photo-production (spectral quantum yield spectrum, AQY for short). In addition, the temperature parameter in the environmental parameter is introduced to fit the photo-production efficiency of the methane.

[0162] It should be noted that the spectral quantum yield spectrum in the embodiment is obtained through a series of experiments related to the photo-production of the methane. For example, the calculation formula is as follows:

[0163]

[0164] where b is a temperature correction factor, and T is temperature.

[0165] Sub-step 2502, based on the diffuse attenuation coefficient and the underwater spectral scalar irradiance, analyze the exponential decay of the underwater spectral scalar irradiance at a specified depth, and combine the absorption coefficient spectrum and the apparent quantum yield to analyze the photon absorption rate at the specified depth.

[0166] In a specific implementation, the spectral scalar irradiance at the surface (z=0-) can be extended to a specified depth z using an exponential decay formula of underwater light field. For example, the formula can be used: The analysis of the exponential decay of the underwater spectral scalar irradiance at a specified depth is implemented to obtain the spectral scalar irradiance E odDay (λ, z).

[0167] where E odDay (λ, 0 - ): the daily integral spectral scalar irradiance under the water surface. K d (λ): the spectral diffuse attenuation coefficient. z: the target calculation depth.

[0168] Then, based on the CDOM absorption coefficient spectrum a g and the irradiance E odDay (λ, z) at the specified depth, the rate of CDOM absorption of photons is calculated. Exemplarily, the calculation of the photon absorption rate can be implemented by the formula: Ξ(λ, z) = E odDay (λ, z) × a g (λ).

[0169] Sub-step 2503, based on the photon absorption rate, perform spectral range integration and vertical integration to obtain the methane photo-production efficiency.

[0170] In this embodiment, after the photon absorption rate is determined, the photon absorption rate can be integrated along the spectral range, and vertically integrated to obtain the methane photo-production efficiency of the mixed layer.

[0171] Exemplarily, the implementation is realized by the following formula:

[0172]

[0173] where MLD: mixed layer depth (generally determined according to external data such as MLD mixed layer depth climatology)

[0174] The unit of the methane photo-production efficiency can be: mol CH4m -2 day -1.

[0175] Sub-step 2504, according to the optical characteristic parameters and the environmental parameter information, the biological production efficiency of methane is inversely analyzed.

[0176] In the embodiment, the biological production efficiency of methane is mainly calculated through the absorption coefficient spectrum and the spectral slope of CDOM, and is fitted and corrected by using the temperature parameter in the environmental parameter, so as to obtain the accurate biological production efficiency of methane.

[0177] Exemplarily, the biological production efficiency can be calculated by the following formula:

[0178]

[0179] Wherein, P bio (a g , S 275-295 , T) is the fitted and corrected biological production efficiency of methane under the current temperature parameter.

[0180] Therefore, the embodiment realizes accurate inversion of the biological production efficiency of methane based on multiple parameter data related to methane production, so as to serve as one of important driving quantities for inversion and prediction of methane concentration.

[0181] Step 260, the in-situ production efficiency, the dissipation information, the methane oxidation rate and the terrestrial input index are subjected to basic time sequence feature analysis and special time sequence feature analysis by a time sequence feature extraction module, so as to obtain first time sequence feature information.

[0182] Wherein, the first time sequence feature information includes short-term dynamic characteristics, average state characteristics, recent state characteristics and long-term trend item characteristics of the sea surface methane concentration.

[0183] Step 270, the environmental parameter information is subjected to special time sequence feature analysis by the time sequence feature extraction module, so as to obtain second time sequence feature information, and the time sequence feature information is generated based on the first time sequence feature information and the second time sequence feature information.

[0184] Wherein, the second time sequence feature information includes relevant features of the sea surface methane concentration change affected by the wind speed.

[0185] The steps 260-270 are uniformly described as follows:

[0186] In a specific implementation, the embodiment fully analyzes the internal dynamic pattern of methane in the "disturbance-accumulation-release" process, deeply studies the process of sea surface methane from production to dispersion, and realizes accurate prediction of sea surface methane concentration and prediction of its dynamic change trend. In order to perfect the inversion and prediction of sea surface methane concentration by the time series neural network model, the embodiment upgrades the structure of the input parameters. Specifically, the embodiment inverses the key driving quantity based on remote sensing data and related parameters, extracts the time series characteristics using the key driving quantity, so that the model can analyze the time series characteristics to obtain the dynamic change prediction result of the sea surface methane concentration.

[0187] In actual implementation, the key driving quantity is combined with the obtained parameter data to perform feature extraction through a time series feature extraction module, including basic time series features and special time series features.

[0188] In extracting the basic time series features, the influence of the driving quantity is analyzed, the short-term dynamic change, the average state change, the recent state change, and the long-term trend change are analyzed, and the special time series process features are analyzed, mainly including the dynamic change of methane, to obtain the first time series features.

[0189] Then, the change of the sea surface methane concentration is affected by the disturbance-recovery mode of the high wind speed event, mainly using environmental parameters to extract special time series process features, including analyzing the influence of wind speed mutation time on methane, the influence of methane accumulation after wind speed mutation, and the influence of wind speed event interval on methane concentration.

[0190] Step 280, based on the in-situ production efficiency, the escape information, the land source input index, and the methane oxidation rate, construct time series data, combine the time series feature information, analyze short-term and long-term features through a preset time series neural network model group, and obtain a first sea surface methane concentration value and a second sea surface methane concentration value predicted in a specified time.

[0191] The time series neural network model group includes a first model and a second model, the first model inverses and analyzes the first sea surface methane concentration value based on the input data, and the second model predicts the second sea surface methane concentration value based on the input data.

[0192] In the embodiment, the models included in the time series neural network model group mainly extract the time series features and the physical driving factors of the methane concentration change (including the obtained parameters and the key driving quantity) in the learning. Through the multi-layer recursive structure, in the model training stage, the short-term and long-term features existing in the input time series are comprehensively learned, the nonlinear mapping relationship between environmental change and methane concentration response is established, so that the inversion of the sea surface methane concentration value of the known target time and the prediction of the sea surface methane concentration value in the future time period can be realized.

[0193] Thus, based on the time series neural network model group, the embodiment deeply studies the internal dynamic pattern of methane in the "disturbance-accumulation-release" process for the production and diffusion of methane at the sea surface position, and realizes the inversion of the sea surface methane concentration value on the one hand and the prediction of the sea surface methane concentration value in the future time period on the other hand.

[0194] Further, in a specific implementation, the embodiment is constructed and trained by in-depth research and analysis to construct and train each time series neural network model (composing the time series neural network model group), a physical quantity derivation module and a time series feature extraction module. In actual application, the physical quantity derivation module can be used to derive the above-mentioned key physical driving quantity according to the wind speed, CDOM, temperature and other environmental variables, and the time series feature extraction module can be used to extract the time series features such as lag term, weighted average, wind speed mutation response and the like from these driving quantities and original inputs. Finally, the long short-term memory neural network (LSTM) model learns the internal dynamic pattern of methane in the "disturbance-accumulation-release" process, so as to realize the accurate inversion and prediction of the sea surface methane concentration value.

[0195] Optionally, before analyzing the short-term and long-term features by the preset time-series neural network model group, the embodiment further includes: obtaining initial samples from a preset database source, the initial samples including sea surface remote sensing reflectance samples, environmental parameter samples, optical parameter samples, diffuse attenuation coefficient samples, downwelling irradiance samples, and photosynthetically active radiation samples; performing standardization processing and time-series data construction processing on the initial samples to obtain time-series data samples with time-series continuity, and dividing the time-series data samples into a training set and a validation set; constructing a time-series feature extraction module and a physical quantity derivation module; training the physical quantity derivation module based on the training set, deriving key driving quantities related to sea surface methane concentration changes, constructing time-series samples, and performing baseline time-series feature and special time-series feature extraction training on the time-series feature extraction module based on the key driving quantities as input to obtain time-series feature samples; constructing an initial first model and an initial second model, and inputting the time-series feature samples and the time-series samples into the initial first model and the initial second model, the initial first model and the initial second model being time-series neural network models and forming an initial time-series neural network model group; the initial time-series neural network model group learns variation characteristics associated with actual physical processes from the input time-series feature samples and time-series samples, and learns short-term features and long-term features from input time-series data, automatically extracts and encodes key information in input time-series through parameter adjustment and iterative optimization, gradually constructs feature representation of time-series for sea surface methane concentration inversion and prediction, and outputs inversion results and prediction results; the validation set is used to verify and evaluate the inversion results and the prediction results respectively, the parameters of the initial time-series neural network model group are optimized until a trained time-series neural network model group is obtained; wherein the initial first model outputs the inversion results, the initial second model outputs the prediction results, the features extracted by each model in the initial time-series neural network model group are black-box learning, which is used to effectively extract feature parameters to improve prediction accuracy, the time-series feature samples include time-series features with clear physical meaning, which are used to enhance model prediction performance and provide scientific interpretability.

[0196] Exemplarily, the present embodiment obtains data samples from data sources, mainly containing three types of independent data sets: ① Global historical research database: MEMENTO (MarinE MethanE and NiTrous Oxide) database (Kock and Bange, 2015), covering shipborne observation data of 26,000 stations worldwide; ② Supplementary data from literature: 12 papers published in journals such as Nature Geoscience (Mao et al., 2022) disclose sea surface methane concentration data after the update period of the MEMENTO database. ③ Independent measurement data (2024-2025), including: X North Sea voyage: 12 stations of methane data obtained by "Experiment No. 6" research ship in January 2024 and 41 stations of methane data obtained by regular research in Beibu Gulf in 2024 and 2025.

[0197] The obtained data samples are sorted to obtain the following key data samples:

[0198] ①, remote sensing reflectivity: from OCCCI (OceanColour-Climate Change Initiative) v4.2; ②, chlorophyll a (Chl-a) and photosynthetically active radiation (PAR): based on OCCCI v6.0 algorithm, combining MODIS / Aqua and VIIRS data, with inversion errors of ±30% and ±5%, respectively; ③, CDOM absorption coefficient and spectral slope; ④, diffuse attenuation coefficient; ⑤, sea surface temperature (SST): NOAA OISST v2.1 multi-sensor fusion product (Huang et al., 2021), spatial resolution 0.25°, absolute accuracy ±0.5℃; ⑥, sea surface wind speed (SSW): NOAA NCEI Blended Seawinds (NBS) v2 product (Saha and Zhang, 2022), combining ASCAT and RapidScat data, 12.5km resolution, error <1.5m / s; ⑦, downward solar irradiance: calculated based on Ruggaber et al. (1994) radiation transfer model, input parameters including solar zenith angle, ozone concentration and aerosol optical thickness.

[0199] Then, the obtained initial data samples are cleaned and standardized, first, the outliers are removed: 3σ principle is used to identify outliers, and values that are obviously abnormal in physics (such as excessively high wind speed) are also identified, then the variables are standardized, the methane concentration data is converted to the same unit, and other input variables are logarithmically converted and standardized according to the specific properties of the parameters to improve the stability of the data distribution.

[0200] Next, spatio-temporal matching and difference, time series data construction are carried out. Among them, spatio-temporal matching: all input data (remote sensing, geophysical model, environmental parameter) need to be matched with the above-mentioned methane expedition database data, and the data source alignment is ensured; spatio-temporal difference: DINEOF empirical orthogonal function interpolation is used to repair the data, and the time series integrity is ensured. Based on the satellite data of the previous 9 days, the day and the next 3 days, the complete time series of the previous 6 days and the day is filled, and for the local missing area, EOF decomposition is used to restore the data continuity.

[0201] Subsequently, the time series data after pretreatment is constructed Mechanism parameter system, specifically, the core mechanism parameters of the model involve methane production (photo production, biological production), consumption (methane oxidation), dissipation (sea-air exchange, bubble flux) and land source input, etc. The physical parameters of these processes are as follows:

[0202] The photo production yield is mainly calculated by using the following formula:

[0203]

[0204] Among them, the first half of the above formula is the number of light quanta absorbed by CDOM in the sea surface mixed layer, which is calculated using satellite remote sensing a g , K d and mixed layer depth (MLD) data. The apparent quantum yield of methane is determined based on the data of methane production experiment.

[0205] The biological production yield is mainly calculated by using the following formula:

[0206] ΔCH4=∫P bio (a g ,S 275-295 ,T,t)dt

[0207] The production efficiency P bio (a g ,S 275-295 ,T,t) of biological production is determined based on the methane production experiment, and changes with the change of the sea surface water environment.

[0208] The methane oxidation rate can be referred to the description of step 150.

[0209] The sea-air exchange flux and bubble flux can be referred to the description of step 240.

[0210] After determining the key driving quantities, the time series neural network model, the physical quantity derivation module and the time series feature extraction module can be constructed.

[0211] When extracting time series characteristic variables (feature engineering), the following is mainly concerned:

[0212] A, basic time series features, including: lag variables: construct lagged values of input parameters for the previous 1 to 7 days to reflect short-term dynamics. Window average: calculate the window average of 2 to 7 days to consider the average state of long-term parameters. Exponential weighted average: weight recent data, taking into account the long-term average state while emphasizing the greater contribution of recent state. Long-term trend: extract the long-term trend item.

[0213] B, special time series process features, including: wind speed sudden event: calculate the wind speed disturbance duration and recovery time. Define a dynamic weighted time window: ensure that the model learns how methane accumulates over time after a wind speed mutation. Wind speed event interval feature: identify the time interval from the last high wind speed event after a high wind speed event, and use it to infer the recovery of methane concentration.

[0214] In the embodiment, the structure of each time series neural network model can be as shown in Figure 5 The model structure can be used to train multiple sub-models to form a "model group" to undertake the tasks of inversion of current concentration and prediction of future concentration. Users can call different models according to task requirements, maintain structural consistency, and enhance adaptability.

[0215] The time series neural network model includes: an input layer for receiving time series of basic environmental driving parameters and extracted time series features with physical meaning. The time series directly comes from remote sensing inversion or geophysical data, reflecting the main physical process of methane change at sea surface. These time series are directly used as input of the neural network and are basic driving information of the model. The time series features (with clear physical meaning) are extracted from the original time series based on the physical understanding of the dynamic process of marine methane. These features serve as additional input channels, which help to improve the physical interpretability of the model and highlight the creative labor of the modeling stage of the scheme. The length of the time series is set to 2-7 days. The long short-term memory (LSTM) sequence processing layer includes: a first layer LSTM to extract short-term dependencies (such as rapid fluctuations and sudden events), and a second layer LSTM to extract long-term trends (such as cumulative effects, lag response, and slow variable evolution). The Dropout layer randomly discards data (p=0.2) to prevent overfitting. The fully connected layer maps the output to the predicted methane concentration. The output layer is mainly used to output the predicted methane concentration.

[0216] Finally, the constructed time series neural network models are trained. In the training process, the LSTM network automatically extracts and encodes the key information in the input time series through a large number of parameter adjustments and iterative optimization, and gradually builds a time series feature representation that can predict the target quantity (sea surface methane concentration).

[0217] Considering that the time series neural network model can perform inversion and prediction tasks, the label time points used during its training can be different. That is, when training the inversion analysis task, samples at the known T0 time point are used; when training the prediction task, samples at the T0+n time point are used.

[0218] During the model training and optimization process, the data is divided into a training set of 90% and a validation set of 10%. The training set adopts ten-fold cross-validation. Hyperparameter optimization is introduced during the training process: the number of neural network units, the learning rate, the Dropout rate, the sequence processing method, the input time series length, and the input parameter. Model verification and independent test: The model verification adopts a three-layer verification framework to ensure the generalization ability of the model, including: internal verification, which uses the global marine methane database, 10% of the data reserved as a test set to evaluate the prediction ability of the model. Independent verification I: independent test using newly published measurement data after 2014 (not involved in training). Independent verification II: verification using self-collected expedition data.

[0219] It should be noted that the time series neural network model constructed in this embodiment can be used for two types of functions: one is an “inversion model”, which inverts the sea surface methane concentration at T0 time under the known input features at the target time (such as T0) and the previous period; the other is a “prediction model”, which only uses the sequence of environmental variables before T0 to predict the concentration at T0+n days in the future under the condition of unknown T0+n related information. The above two types of models are consistent in structure design and input feature selection, and only differ in the training target variable time position, so they can be classified as a group of model ensembles. By adjusting the label time position of the training set, the inversion or prediction function can be flexibly realized.

[0220] After completing the model construction and training in this embodiment, the trained model can be used to form a time series neural network model group to invert the known time sea surface methane concentration value and predict the future time period sea surface methane concentration value from the input time series feature data.

[0221] In summary, the embodiment of the present application uses the optical parameters and the diffuse attenuation coefficient obtained by the inversion analysis of the remote sensing reflectance of the sea surface, combined with environmental parameters, methane concentration, methane oxidation efficiency, etc., to inversely analyze the sea-air exchange flux, bubble flux, land source input index, and other key driving quantities related to the change process of the sea surface methane concentration, and form time series data. Subsequently, the key driving quantities are used to extract time series features with clear physical meaning, and the time series data are jointly input into a time series neural network model group. The key information in the input time series is automatically extracted and coded by the time series neural network model group to identify short-term and long-term features in the time series. On the one hand, the sea surface methane concentration value at a known target time is accurately inverted, and on the other hand, the methane concentration value in a future time period is predicted, thereby improving the monitoring efficiency and applicability.

[0222] As can be seen, the embodiment of the present application deeply studies and analyzes the influence of optical properties, environmental parameters, and various driving quantities on the production, emission, and concentration change of the sea surface methane, thereby achieving accurate inversion and prediction of the sea surface methane concentration value. The present embodiment provides theoretical support for the study of methane production and emission, and achieves accurate inversion calculation of in-situ methane production. The present embodiment quantifies the in-situ methane production efficiency through a physical model, thereby significantly improving the credibility of the inversion and prediction results of the methane concentration.

[0223] The technical solution of the present embodiment solves the problem that the prior art does not pay close attention to various factors related to the change process of methane production and emission, thereby failing to dynamically and accurately invert the sea surface methane concentration value and emission trend.

[0224] Further, the present application achieves a breakthrough in the problems of spatial limitations, time lags, and parameter omissions in the monitoring of the sea surface water body methane concentration on the basis of the prior art. Specifically, the present application has the following advantages / technical effects: ① high time resolution and large-scale sea surface methane concentration monitoring are achieved, which can reduce the observation cost and is efficient compared to the prior art; ② the in-situ methane production efficiency, time series information, and key driving quantities are incorporated, and the large-scale, continuous, and high time resolution methane concentration inversion is achieved through the combination of prediction and time series modeling, thereby making up for the spatial and temporal limitations of traditional methods and effectively improving the inversion accuracy; ③ the dynamic monitoring and long-term trend analysis requirements are met, and high time resolution and spatial continuity are achieved; ④ the present application can provide strong data support for global ocean methane cycle research, climate change assessment, and environmental protection, and has a broad application prospect and significant technical advantages.

[0225] It should be noted that, for the method embodiment, in order to simply describe, it is expressed as a combination of a series of actions, but those skilled in the art should know that the present embodiment is not limited by the order of the described actions, because according to the present embodiment, certain steps can be performed in other order or simultaneously.

[0226] As Figure 6 shown, the embodiment of the present application also provides a sea surface methane concentration inversion and prediction system 600 based on a time series neural network, comprising:

[0227] A data acquisition module 610 is configured to acquire water color remote sensing data and environmental parameter information of a target area within a preset time period, wherein the water color remote sensing data comprises sea surface remote sensing reflectivity acquired by a satellite.

[0228] An optical inversion module 620 is configured to perform water body optical inversion based on the sea surface remote sensing reflectivity to obtain optical characteristic parameters of colored soluble organic matter and a diffuse attenuation coefficient of the water body, wherein the optical characteristic parameters comprise an absorption coefficient spectrum and a spectral slope at a key wavelength.

[0229] A dissipation and oxidation rate analysis module 630 is configured to perform analysis and processing based on the environmental parameter information and acquired methane concentration information to obtain dissipation information and a methane oxidation rate, wherein the dissipation information comprises a sea-air exchange flux and a bubble flux.

[0230] An in-situ production efficiency analysis module 640 is configured to analyze a terrestrial input index according to the spectral slope, and to analyze in-situ production efficiency of sea surface methane according to the optical characteristic parameters, the diffuse attenuation coefficient and the environmental parameter information, wherein the in-situ production efficiency comprises a photo production efficiency and a biological production efficiency.

[0231] A feature extraction module 650 is configured to take the in-situ production efficiency, the dissipation information, the terrestrial input index, the methane oxidation rate and the environmental parameter information as inputs, to perform feature extraction of basic time series and special time series through a time series feature extraction module, and to obtain time series feature information.

[0232] A sea surface methane concentration inversion and prediction module 660 is configured to construct time series data based on the in-situ production efficiency, the dissipation information, the terrestrial input index and the methane oxidation rate, to combine the time series feature information, to analyze short-term and long-term features through a preset time series neural network model group, and to obtain a first sea surface methane concentration value and a second sea surface methane concentration value predicted within a specified time.

[0233] The time sequence neural network model group includes a first model and a second model, the first model is used to inversely analyze the first sea surface methane concentration value based on input data, the second model is used to predict the second sea surface methane concentration value based on input data, the initial acquisition of the methane concentration information is obtained by inversely analyzing based on climatological statistics, and is continuously updated during the running of the time sequence neural network model group, and the initial acquisition of the methane oxidation rate and the dissipation information is obtained by inversely analyzing based on initial methane concentration information, and is continuously updated during the running of the time sequence neural network model group.

[0234] It should be noted that the sea surface methane concentration inversion and prediction system based on the time sequence neural network provided in the embodiments of the present application can perform the sea surface methane concentration inversion and prediction method based on the time sequence neural network provided in any of the embodiments of the present application, and has the functions and advantages corresponding to the execution method.

[0235] In a specific implementation, the above-mentioned sea surface methane concentration inversion and prediction system based on the time sequence neural network can be integrated in a device, so that the device can use the sea surface remote sensing reflectivity to inversely analyze the key driving quantity, extract the time sequence characteristics, and predict the time sequence characteristics through the time sequence neural network model to output the predicted sea surface concentration value, as an electronic device, to realize dynamic and accurate inversion of the sea surface water body methane concentration and concentration change. The electronic device can be composed of two or more physical entities, or can be composed of one physical entity, for example, the electronic device can be a personal computer (PC), a computer, a server, etc., and the embodiments of the present application do not make specific limitations thereto.

[0236] The embodiments of the present application also provide a computer readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the steps of the sea surface methane concentration inversion and prediction method based on the time sequence neural network provided in any one of the preceding method embodiments.

[0237] It should be noted that in this document, relational terms such as "first" and "second", and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the phrase "comprising a... " does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0238] The foregoing detailed description of the application has been presented for purposes of illustration and description. Various modifications and changes can be made to the embodiments in light of the above detailed description without departing from the spirit and intended scope of the application. It is to be understood that the application can be practiced otherwise than as specifically described without altering its spirit or essential characteristics. Accordingly, the scope of the application should be judged in terms of the broadest allowable principles and features described herein, and not according to the above detailed description, which is to be viewed in its more limited sense.

Claims

1. A method for predicting sea surface methane concentration based on time-series neural networks, characterized in that, include: Acquire water color remote sensing data and environmental parameter information of the target area within a preset time period, wherein the water color remote sensing data includes sea surface remote sensing reflectance acquired by satellite; Based on the sea surface remote sensing reflectance, water body optical inversion is performed to obtain the optical characteristic parameters of colored soluble organic matter and the diffuse attenuation coefficient of the water body. The optical characteristic parameters include the absorption coefficient spectrum and spectral slope at key wavelengths. Based on the environmental parameter information and the obtained methane concentration information, analysis and processing are performed to obtain emission information and methane oxidation rate. The emission information includes air-sea exchange flux and bubble flux. The terrestrial input index is analyzed based on the spectral slope, and the in-situ production efficiency of sea surface methane is analyzed by inversion based on the optical characteristic parameters, the diffuse attenuation coefficient and the environmental parameter information. The in-situ production efficiency includes photoproduction efficiency and biological production efficiency. Using the in-situ production efficiency, the dissipation information, the land-based input indicators, the methane oxidation rate, and the environmental parameter information as inputs, the time-series feature extraction module performs feature extraction on the basic time series and special time series to obtain time-series feature information; Based on the in-situ production efficiency, the dissipation information, the land-based input indicators, and the methane oxidation rate, time series data is constructed. Combined with the time series feature information, short-term and long-term features are analyzed through a preset time series neural network model group to obtain the first sea surface methane concentration value and the second sea surface methane concentration value predicted within a specified time. The time-series neural network model group includes a first model and a second model. The first model is based on the input data to invert and analyze the first sea surface methane concentration value, and the second model is based on the input data to predict the second sea surface methane concentration value. The initial acquisition of the methane concentration information is based on the inversion analysis of climatological statistics, and is continuously updated during the operation of the time-series neural network model group. The initial acquisition of the methane oxidation rate and the emission information is based on the initial methane concentration information to invert and analyze, and is continuously updated during the operation of the time-series neural network model group.

2. The method according to claim 1, characterized in that, Based on the sea surface remote sensing reflectance, water body optical inversion is performed to obtain the optical characteristic parameters of colored soluble organic matter and the diffuse attenuation coefficient of the water body, including: Principal component analysis and color cluster classification were performed based on the sea surface remote sensing reflectance, and multiple linear regression calculations were conducted to obtain the absorption coefficient spectrum and spectral slope of colored soluble organic matter. Principal component analysis and cluster analysis were performed on the sea surface remote sensing reflectance, and the water body diffuse attenuation coefficient was obtained by optimization through an optimization algorithm.

3. The method according to claim 2, characterized in that, Principal component analysis and color cluster classification were performed based on the sea surface remote sensing reflectance, and multiple linear regression calculations were conducted to obtain the absorption coefficient spectra and spectral slopes of colored soluble organic matter, including: according to The sea surface remote sensing reflectance was logarithmically transformed, and based on... The logarithmically transformed sea surface remote sensing reflectance is standardized to obtain the first remote sensing reflectance. Input the first remote sensing reflectance into the pre-trained feature vector matrix, according to [PC1 i PC2 i ,PC3 i ]=R′ rs (λ) i ×U performs principal component analysis and calculates the values ​​of the three principal components; The classification of ocean color clusters is based on the three principal component values, and the target cluster number to which each principal component value belongs in the ocean color cluster is determined. Using the target cluster number as a benchmark, the multiple regression linear equation is invoked, based on ln(a g (λ))=β0(λ)+β1(λ)*PC1 i +β2(λ)*PC2 i +β3(λ)*PC3 i Calculate the absorption coefficient of the sea surface remote sensing reflectance at each wavelength, and calculate the complete absorption coefficient spectrum a. g (λ); Using the first remote sensing reflectance at each wavelength as input, according to ln(S) 275-295 )=α+β*ln[R rs (443)]+γ*ln[R rs (488)]+δ*ln[R rs (531)]+ε*ln[R rs (555)]+ζ*ln[R rs (667)], Analyze and calculate the spectral slope S of interest. 275-295 ; Among them, R′ rs (λ) i Let be the first remote sensing reflectance, U be the feature vector matrix obtained through algorithm training, PC1, PC2 and PC3 are principal component values, α, β, γ, δ, ε and ζ are regression coefficients, and mean and std are parameters predetermined in the algorithm.

4. The method according to claim 2, characterized in that, Principal component analysis and cluster analysis were performed on the sea surface remote sensing reflectance, and optimization was carried out using an optimization algorithm to obtain the diffuse attenuation coefficient of the water body, including: The satellite bands corresponding to the sea surface remote sensing reflectance are analyzed, and the satellite bands are standardized to obtain the second remote sensing reflectance belonging to the preset wavelength. The water body type is determined based on the second remote sensing reflectance, and an optimized algorithm is selected based on the water body type determination result; according to The second remote sensing reflectance is normalized to obtain normalized data X(λ); The normalized data X(λ) for each wavelength, combined with the feature vector of the optimization algorithm, are used as input, according to PC. j =e j1 X(412)+e j2 X(443)+e j3 X(490)+e j4 X(510)+e j5 X(555)+e j6 X(670), analyze and calculate the values ​​of the four principal components; The four principal component values ​​are input into the multiple linear regression equation model of the optimization algorithm, according to ln[K d [(λ)]=α+βPC1+γPC2+δPC3+∈PC4, calculate the wavelength diffusion attenuation coefficients corresponding to the wavelengths of the four principal component values; according to By taking the exponent of each wavelength's diffuse attenuation coefficient, the true diffuse attenuation coefficient K is obtained. d (λ); Where λ is the data wavelength, σ R(λ) The standard deviation is predetermined based on the water body type. e is an average value predetermined based on the type of water body. j Given the feature vector.

5. The method according to claim 1, characterized in that, Based on the environmental parameter information and the acquired methane concentration information, analysis and processing are performed to obtain emission information and methane oxidation rate, including: Using the temperature, salinity, and water depth information from the environmental parameters and the methane concentration information as inputs, a preset random forest regression model is used to invert the methane oxidation rate; Using the wind speed information from the environmental parameter information as input, according to Calculate the gas transport rate k; Using the gas transport rate k as input, according to F diff =k([CH4]) water -[CH4] air ), calculate the air-sea exchange flux of methane F diff ; according to Calculate the bubble flux; in, Here, is the wind speed above the sea surface, Sc is the Schmidt number, and [CH4] represents the wind speed. water The concentration of methane in sea surface waters, [CH4]. air This represents the methane concentration above the sea surface.

6. The method according to claim 1, characterized in that, The analysis of land-source input indicators based on the spectral slope includes: according to Analyze the characteristic information of the spectral slope S; Using the aforementioned characteristic information as input, according to f 陆源 =[ln(S) 样品 )-ln(S 陆源端元 )] / [ln(S 海洋端元 )-ln(S 陆源端元 )], quantitatively calculate the land-source input index f_landsource; The characteristic information includes the features of gently decreasing absorption curves and steeply decreasing absorption curves, as well as the absorption curve features of the sample with the gentlest absorption curve decrease and the smallest spectral slope in the sample space, characterizing S. 陆源端元 The absorption curve characteristics of the sample with the steepest descent and the largest spectral slope in the sample space, characterizing S 海洋端元 .

7. The method according to claim 1, characterized in that, Based on the optical characteristic parameters, the diffuse attenuation coefficient, and the environmental parameter information, the in-situ production efficiency of sea surface methane is analyzed, including: Based on the environmental parameter information and the absorption coefficient spectrum, the spectral quantum yield spectrum related to methane photoproduction is inverted and analyzed, and the underwater spectral scalar irradiance is obtained through a radiative transfer model. Based on the diffuse attenuation coefficient and the underwater spectral scalar irradiance, the exponential decay of the underwater spectral scalar irradiance at a specified depth is analyzed, and combined with the absorption coefficient spectrum and apparent quantum yield, the photon absorption rate at a specified depth is analyzed. Based on the photon absorption rate, the spectral range integration and vertical integration are performed to obtain the methane photo-production efficiency. Based on the optical characteristic parameters and environmental parameter information, the bioproduction efficiency of methane was analyzed by inversion.

8. The method according to claim 1, characterized in that, Using the in-situ production efficiency, the emission information, the terrestrial input indicators, the methane oxidation rate, and the environmental parameter information as inputs, a time-series feature extraction module is used to extract features from the basic time series and specific time series to obtain time-series feature information, including: The in-situ production efficiency, the emission information, the methane oxidation rate, and the land-based input indicators are analyzed using a time-series feature extraction module to obtain the first time-series feature information. The environmental parameter information is analyzed by the time-series feature extraction module to obtain the second time-series feature information, and time-series feature information is generated based on the first time-series feature information and the second time-series feature information. The first time-series feature information includes short-term dynamic features, average state features, recent state features, and long-term trend features of sea surface methane concentration, while the second time-series feature information includes relevant features of sea surface methane concentration changes affected by wind speed.

9. The method according to any one of claims 1-8, characterized in that, Before analyzing short-term and long-term features using a pre-defined set of time-series neural network models, the following steps are also included: Initial samples are obtained from a preset database source, including sea surface remote sensing reflectance samples, environmental parameter samples, optical parameter samples, diffuse attenuation coefficient samples, downlink irradiance samples, and photosynthetically active radiation samples. Based on the initial samples, standardization and time series data construction processes are performed to obtain time series data samples with temporal continuity, and the time series data samples are divided into training set and validation set; Construct a time-series feature extraction module and a physical quantity derivation module; Using the training set as a benchmark, the physical quantity derivation module is trained to derive key driving quantities related to changes in sea surface methane concentration, construct time series samples, and use the key driving quantities as input to train the time series feature extraction module to extract benchmark time series features and special time series features, thereby obtaining time series feature samples. An initial first model and an initial second model are constructed, and the time-series feature samples and the time series samples are used as inputs. Both the initial first model and the initial second model are time-series neural network models, and they form an initial time-series neural network model group. The initial time-series neural network model group learns the changing characteristics associated with actual physical processes from the input time-series feature samples and time series samples, as well as short-term and long-term features from the input time series data. Through parameter adjustment and iterative optimization, it automatically extracts and encodes key information in the input time series, gradually constructs a feature representation of the time series used for inversion and prediction of sea surface methane concentration, and outputs inversion results and prediction results. The inversion results and the prediction results are verified and evaluated using the validation set, and the parameters of the initial temporal neural network model group are optimized until a well-trained temporal neural network model group is obtained. The initial first model outputs the inversion result, the initial second model outputs the prediction result, the features extracted from each model in the initial temporal neural network model group are learned in a black box manner, which is used to effectively extract feature parameters to improve prediction accuracy, and the temporal feature samples include temporal features with clear physical meaning, which are used to enhance the model prediction performance and provide scientific interpretation.

10. A sea surface methane concentration inversion and prediction system based on time series neural networks, characterized in that, include: The data acquisition module is used to acquire water color remote sensing data and environmental parameter information of the target area within a preset time period. The water color remote sensing data includes sea surface remote sensing reflectance acquired by satellite. The optical inversion module is used to perform optical inversion of water bodies based on the sea surface remote sensing reflectance to obtain the optical characteristic parameters of colored soluble organic matter and the diffuse attenuation coefficient of the water body. The optical characteristic parameters include the absorption coefficient spectrum and spectral slope at key wavelengths. The escape and oxidation rate analysis module is used to analyze and process the environmental parameter information and the acquired methane concentration information to obtain escape information and methane oxidation rate. The escape information includes air-sea exchange flux and bubble flux. The in-situ production efficiency analysis module is used to analyze land-based input indicators based on the spectral slope, and to invert and analyze the in-situ production efficiency of sea surface methane based on the optical characteristic parameters, the diffuse attenuation coefficient and the environmental parameter information. The in-situ production efficiency includes photoproductivity and bioproductivity. The feature extraction module is used to extract basic time series and special time series features by taking the in-situ production efficiency, the dissipation information, the land-based input index, the methane oxidation rate and the environmental parameter information as input, and obtaining time series feature information by using the time series feature extraction module. The sea surface methane concentration inversion and prediction module is used to construct time series data based on the in-situ production efficiency, the emission information, the land-based input indicators and the methane oxidation rate, and combine the time series feature information to analyze short-term and long-term features through a preset time series neural network model group to obtain the first sea surface methane concentration value and the second sea surface methane concentration value predicted within a specified time. The time-series neural network model group includes a first model and a second model. The first model is based on the input data to invert and analyze the first sea surface methane concentration value, and the second model is based on the input data to predict the second sea surface methane concentration value. The initial acquisition of the methane concentration information is based on the inversion analysis of climatological statistics, and is continuously updated during the operation of the time-series neural network model group. The initial acquisition of the methane oxidation rate and the emission information is based on the initial methane concentration information to invert and analyze, and is continuously updated during the operation of the time-series neural network model group.

Images