Method for reconstructing ocean chlorophyll and light field vertical profile driven by sea surface

Abstract

The application discloses a sea surface driving type method for reconstructing a vertical profile of marine chlorophyll and a light field, belongs to the technical field of marine remote sensing and marine monitoring, and is used for vertical profile reconstruction and comprises the following steps: screening an Argo profile, extracting surface photosynthetically active radiation, surface chlorophyll concentration and sea surface temperature, calculating three-dimensional unit spherical coordinates and observation time periodic coding of an observation point, and constructing an input vector; a deep learning model of double-branch joint inversion is constructed and trained, an Argo profile to be predicted is input into the trained deep learning model of double-branch joint inversion, and a final chlorophyll concentration profile prediction result is output. In the chlorophyll profile inversion process, the application explicitly introduces sea surface and profile photosynthetically active radiation information, so that the model can more truly reflect the influence of light conditions in the water body on the vertical distribution of chlorophyll, effectively reduces inversion uncertainty, and improves the overall inversion effect.

Description

Technical Field

[0001] This invention discloses a sea surface-driven method for reconstructing the vertical profile of marine chlorophyll and light field, belonging to the field of marine remote sensing and marine monitoring technology. Background Technology

[0002] Ocean remote sensing is a crucial means of acquiring large-scale, long-term ocean information. Traditional satellite observations primarily acquire sea surface parameters; however, most biogeochemical processes in the ocean occur within the water body, especially vertically. Therefore, retrieving vertical profile information of the water body using sea surface remote sensing signals has become a core challenge in this field. Among these, photosynthetically active radiation profiles and chlorophyll concentration profiles are two key parameters for understanding marine primary productivity and ecosystem structure, and achieving accurate and coordinated retrieval of these parameters is of significant scientific importance.

[0003] Currently, there are two main but independent technical approaches for the inversion of these two types of parameters. One is a semi-analytical algorithm-based method for photosynthetically active radiation (PAR) profile inversion. This method uses the physical relationship between the sea surface light field and the vertical distribution of chlorophyll to decompose PAR using empirical formulas and simulate its attenuation process in the water body, thereby achieving profile estimation. The other is a deep learning-based chlorophyll profile inversion method. This method uses sea surface remote sensing data and measured profile data to construct a training set, and directly learns the complex mapping relationship from sea surface features to vertical concentration distribution through a neural network, achieving end-to-end profile prediction.

[0004] The aforementioned existing technologies have significant limitations: First, the inversion of photosynthetically active radiation relies on fixed physical formulas and empirical assumptions, making it ill-suited for optically complex water environments and unable to learn and optimize autonomously from data. Second, the two inversion techniques are completely separate. When both types of profile data need to be used simultaneously, they originate from physical models and data-driven models respectively, lacking necessary physical constraints that lead to inherent data inconsistencies and affect the reliability of subsequent scientific analysis. Furthermore, existing deep learning models only invert single parameters, failing to consider the interaction between the light field and pigment distribution during the learning process, thus limiting the model's physical rationality and its potential for joint inversion. Summary of the Invention

[0005] The purpose of this invention is to provide a sea surface-driven method for reconstructing marine chlorophyll and light field vertical profiles, in order to solve the problems in the prior art, which rely on traditional semi-analysis algorithms, lack specialized and efficient deep learning models, fail to coordinate the chlorophyll concentration profile inversion model with the light field profile for learning and constraint, and have independent and poor physical coordination between the light field profile and the chlorophyll concentration profile inversion.

[0006] Sea surface-driven methods for reconstructing marine chlorophyll and light field vertical profiles include:

[0007] S1. Set filtering conditions to filter Argo profiles. Argo profiles include chlorophyll concentration profiles, ocean temperature profiles, and photosynthetically active radiation profiles. Standardize the filtered Argo profiles and perform physical fitting to generate surface photosynthetically active radiation and surface chlorophyll concentration. Convert the latitude and longitude of the observation points to three-dimensional unit spherical coordinates. Extract the date and time information of the observation points within the year to construct a periodic code. Extract the sea surface temperature of the Argo profile. Linearly normalize the sea surface temperature, surface chlorophyll concentration, and surface photosynthetically active radiation to construct the input vector.

[0008] S2. Construct a deep learning model for dual-branch joint inversion, train the neural network, take the input vector as input, and take the real photosynthetically active radiation profile and the real chlorophyll concentration profile as real labels. Constrain the deep learning model for dual-branch joint inversion through a segregated loss function, update the dual-branch training parameters through a step-by-step backpropagation mechanism, set an iteration number threshold, and output the trained dual-branch joint inversion deep learning model when the iteration number reaches the iteration number threshold.

[0009] After the Argo profile to be predicted is processed in step S1, it is input into the trained dual-branch joint inversion deep learning model, and the final chlorophyll concentration profile prediction result is output.

[0010] S1 includes, S1.1, setting the photosynthetically effective radiation depth threshold. , obtain Photosynthetically active radiation to the surface Calculate the effective photosynthetic radiation of the surface layer using a quadratic function. :

[0011] ;

[0012] ;

[0013] In the formula, , , The coefficients are quadratic polynomials. It is a natural exponential function;

[0014] S1 includes S1.2, and the depth of the photosynthetically active radiation profile is... ,Will Linear interpolation to ,get And perform a logarithmic transformation:

[0015] ;

[0016] In the formula, Photosynthetically active radiation label;

[0017] S1 includes, S1.3, performing point smoothing on the chlorophyll concentration profile, assuming the chlorophyll concentration profile depth is... Set chlorophyll concentration depth threshold , obtain chlorophyll concentration to the surface The chlorophyll concentration profile was characterized using a Gaussian parametric model:

[0018] ;

[0019] In the formula, These are chlorophyll concentration profile parameters. , This represents the background concentration of the chlorophyll concentration profile. The thickness of the subsurface chlorophyll layer is the maximum value. This represents the depth corresponding to the maximum chlorophyll concentration. The parameter is used to control the amplitude of the peak term in the vertical distribution of chlorophyll. Indicates depth as The unit is meters. For depth indexing, ;

[0020] S1 includes, S1.4, and... Perform initialization, including Initialize to the minimum value of the chlorophyll concentration profile. Initialized as the difference between the maximum and minimum values ​​of the chlorophyll concentration profile. Set a fixed initial value. The depth corresponding to the maximum chlorophyll concentration;

[0021] Use the first linear least squares method to initialize the... Estimate the optimal chlorophyll concentration profile parameters to obtain the optimal parameters. :

[0022] ;

[0023] In the formula, for Observed chlorophyll concentration at [location];

[0024] S1 includes, S1.5, from 0 to Depth sequence of rules constructed within the range , For chlorophyll depth index, ,Will Substitution ,calculate chlorophyll concentration value, Indicates depth as The unit is meters.

[0025] S1 includes S1.6, converting the latitude and longitude of Argo profile observation points into three-dimensional unit spherical coordinates:

[0026] ;

[0027] ;

[0028] ;

[0029] In the formula, The latitude of the observation point. The longitude of the observation point. The x-coordinate of the observation point. The y-coordinate of the observation point. The z-axis coordinates of the observation point;

[0030] S1 includes S1.7, extracting the date and time information within the year from the observation time of the observation point, and using sine and cosine functions for periodic encoding:

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] In the formula, This refers to the number of days in the year. For a moment, The sine function is used to periodically encode the dates within the year. The periodic encoding of dates within the year using the cosine function. The sinusoidal function is periodically encoded to represent the time information. The periodic encoding of time information is a cosine function.

[0036] S1 includes, S1.8, setting the sea surface temperature depth threshold. , obtain Sea surface temperature The linear normalization formula is constructed as follows:

[0037] ;

[0038] In the formula, These are the original values ​​of the corresponding variables. The minimum value of the corresponding data. The maximum value of the corresponding data. The normalized corresponding variables include... , and .

[0039] S1 includes, S1.9, using the linear normalization formula to... , and Linear normalization was performed to obtain the normalized sea surface temperature. Normalized sea surface chlorophyll concentration and normalized sea surface photosynthetically active radiation Construct the input vector :

[0040] .

[0041] S2 includes constructing a deep learning model ZS-Net with dual-branch joint inversion. ZS-Net consists of two branches: the first branch adopts a multilayer perceptron structure, including a feature extraction module, an output layer, and a dimension expansion layer; the second branch adopts a hybrid structure of multilayer perceptron and bidirectional LSTM, including a feature extraction module, a feature fusion layer, a sequence modeling module, and an output layer.

[0042] The feature extraction module of the first branch includes three fully connected layers. Each fully connected layer includes a Linear layer and a Tanh activation function. The output layer is a Linear layer, and the dimension expansion layer is a Linear layer.

[0043] The feature extraction module of the second branch includes three fully connected layers. Each fully connected layer includes a Linear layer and a Tanh activation function. The feature fusion layer is a concat layer. The sequence modeling module includes a forward LSTM layer and a backward LSTM layer. The output layer is a Linear layer.

[0044] The output of the first branch is input into the feature fusion layer of the second branch, and then concatenated in the last dimension by the concat layer to form a sequence input. , For batches, The depth of the chlorophyll profile is the number of layers; The sequence modeling module outputs 32-dimensional features from the forward LSTM layer and the backward LSTM layer, which are then concatenated to obtain 64-dimensional features. The results of the sequence modeling module are then input into the output layer of the second branch to output the chlorophyll concentration profile prediction results.

[0045] Using true photosynthetically active radiation and true chlorophyll concentration as true values, a segregated loss function is constructed, which includes a first branch loss. Second branch loss , This represents the mean square error of photosynthetically active radiation. This represents the mean square error of chlorophyll concentration.

[0046] A step-by-step backpropagation mechanism is adopted. Update the training parameters for the first branch. Update the training parameters for the second branch; the second branch uses a detach operation to prevent... Backpropagation to the first branch.

[0047] Compared with the prior art, the present invention has the following beneficial effects: The present invention explicitly introduces the photosynthetically active radiation information of the sea surface and the profile during the chlorophyll profile inversion process, so that the model can more realistically reflect the influence of light conditions on the vertical distribution of chlorophyll in the water body, effectively reduce the uncertainty of inversion, and improve the overall inversion effect. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the deep learning model structure for dual-branch joint inversion of the present invention;

[0049] Figure 2 This is the prediction result of the CHL profile on the BGC-Argo test set (Part 1);

[0050] Figure 3 This is the prediction result of the CHL profile on the BGC-Argo test set (Part 2);

[0051] Figure 4 This is the PAR profile prediction result for the BGC-Argo test set (Part 1);

[0052] Figure 5 This is the prediction result of the PAR profile on the BGC-Argo test set (II).

[0053] Figure 6 This is the prediction result of the CHL profile on the MODIS test set (Part 1).

[0054] Figure 7 This is the prediction result of the CHL profile on the MODIS test set (II).

[0055] Figure 8 This is the MODIS test set PAR profile prediction result (I);

[0056] Figure 9 This is the MODIS test set PAR profile prediction result (II). Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0058] Sea surface-driven methods for reconstructing marine chlorophyll and light field vertical profiles include:

[0059] S1. Set filtering conditions to filter Argo profiles. Argo profiles include chlorophyll concentration profiles, ocean temperature profiles, and photosynthetically active radiation profiles. Standardize the filtered Argo profiles and perform physical fitting to generate surface photosynthetically active radiation and surface chlorophyll concentration. Convert the latitude and longitude of the observation points to three-dimensional unit spherical coordinates. Extract the date and time information of the observation points within the year to construct a periodic code. Extract the sea surface temperature of the Argo profile. Linearly normalize the sea surface temperature, surface chlorophyll concentration, and surface photosynthetically active radiation to construct the input vector.

[0060] S2. Construct a deep learning model for dual-branch joint inversion, train the neural network, take the input vector as input, and take the real photosynthetically active radiation profile and the real chlorophyll concentration profile as real labels. Constrain the deep learning model for dual-branch joint inversion through a segregated loss function, update the dual-branch training parameters through a step-by-step backpropagation mechanism, set an iteration number threshold, and output the trained dual-branch joint inversion deep learning model when the iteration number reaches the iteration number threshold.

[0061] After the Argo profile to be predicted is processed in step S1, it is input into the trained dual-branch joint inversion deep learning model, and the final chlorophyll concentration profile prediction result is output.

[0062] S1 includes, S1.1, setting the photosynthetically effective radiation depth threshold. , obtain Photosynthetically active radiation to the surface Calculate the effective photosynthetic radiation of the surface layer using a quadratic function. :

[0063] ;

[0064] ;

[0065] In the formula, , , The coefficients are quadratic polynomials. It is a natural exponential function;

[0066] S1 includes S1.2, and the depth of the photosynthetically active radiation profile is... ,Will Linear interpolation to ,get And perform a logarithmic transformation:

[0067] ;

[0068] In the formula, Photosynthetically active radiation label;

[0069] S1 includes S1.3, performing point smoothing on the chlorophyll concentration profile, assuming the chlorophyll concentration profile depth is... Set chlorophyll concentration depth threshold , obtain chlorophyll concentration to the surface The chlorophyll concentration profile was characterized using a Gaussian parametric model:

[0070] ;

[0071] In the formula, These are chlorophyll concentration profile parameters. , This represents the background concentration of the chlorophyll concentration profile. The thickness of the subsurface chlorophyll layer is the maximum value. This represents the depth corresponding to the maximum chlorophyll concentration. The parameter is used to control the amplitude of the peak term in the vertical distribution of chlorophyll. Indicates depth as The unit is meters. For depth indexing, ;

[0072] S1 includes, S1.4, and... Perform initialization, including Initialize to the minimum value of the chlorophyll concentration profile. Initialized as the difference between the maximum and minimum values ​​of the chlorophyll concentration profile. Set a fixed initial value. The depth corresponding to the maximum chlorophyll concentration;

[0073] Use the first linear least squares method to initialize the... Estimate the optimal chlorophyll concentration profile parameters to obtain the optimal parameters. :

[0074] ;

[0075] In the formula, for Observed chlorophyll concentration at [location];

[0076] S1 includes, S1.5, from 0 to Depth sequence of rules constructed within the range , For chlorophyll depth index, ,Will Substitution ,calculate chlorophyll concentration value, Indicates depth as The unit is meters.

[0077] S1 includes S1.6, converting the latitude and longitude of Argo profile observation points into three-dimensional unit spherical coordinates:

[0078] ;

[0079] ;

[0080] ;

[0081] In the formula, The latitude of the observation point. The longitude of the observation point. The x-coordinate of the observation point. The y-coordinate of the observation point. The z-axis coordinates of the observation point;

[0082] S1 includes S1.7, extracting the date and time information within the year from the observation time of the observation point, and using sine and cosine functions for periodic encoding:

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] In the formula, This refers to the number of days in the year. For a moment, The sine function is used to periodically encode the dates within the year. The periodic encoding of dates within the year using the cosine function. The sinusoidal function is periodically encoded to represent the time information. The periodic encoding of time information is a cosine function.

[0088] S1 includes, S1.8, setting sea surface temperature depth thresholds. , obtain Sea surface temperature The linear normalization formula is constructed as follows:

[0089] ;

[0090] In the formula, These are the original values ​​of the corresponding variables. The minimum value of the corresponding data. The maximum value of the corresponding data. The normalized corresponding variables include... , and .

[0091] S1 includes, S1.9, using the linear normalization formula to... , and Linear normalization was performed to obtain the normalized sea surface temperature. Normalized sea surface chlorophyll concentration and normalized sea surface photosynthetically active radiation Construct the input vector :

[0092] .

[0093] S2 includes constructing a deep learning model ZS-Net with dual-branch joint inversion. ZS-Net consists of two branches: the first branch adopts a multilayer perceptron structure, including a feature extraction module, an output layer, and a dimension expansion layer; the second branch adopts a hybrid structure of multilayer perceptron and bidirectional LSTM, including a feature extraction module, a feature fusion layer, a sequence modeling module, and an output layer.

[0094] The feature extraction module of the first branch includes three fully connected layers. Each fully connected layer includes a Linear layer and a Tanh activation function. The output layer is a Linear layer, and the dimension expansion layer is a Linear layer.

[0095] The feature extraction module of the second branch includes three fully connected layers. Each fully connected layer includes a Linear layer and a Tanh activation function. The feature fusion layer is a concat layer. The sequence modeling module includes a forward LSTM layer and a backward LSTM layer. The output layer is a Linear layer.

[0096] The output of the first branch is input into the feature fusion layer of the second branch, and then concatenated in the last dimension by the concat layer to form the sequence input. , For batches, The number of chlorophyll profile depth layers; The sequence modeling module outputs 32-dimensional features from the forward LSTM layer and the backward LSTM layer, which are then concatenated to obtain 64-dimensional features. The results of the sequence modeling module are then input into the output layer of the second branch to output the chlorophyll concentration profile prediction results.

[0097] Using true photosynthetically active radiation and true chlorophyll concentration as true values, a segregated loss function is constructed, which includes a first branch loss. Second branch loss , This represents the mean square error of photosynthetically active radiation. This represents the mean square error of chlorophyll concentration.

[0098] A step-by-step backpropagation mechanism is adopted. Update the training parameters for the first branch. Update the training parameters for the second branch; the second branch uses a detach operation to prevent... Backpropagation to the first branch.

[0099] The following description, in conjunction with the accompanying drawings and embodiments, provides further details. Data was obtained from BGC-Argo buoy profiles. Argo profiles containing chlorophyll concentration (CHL), sea temperature (Sea_Temperature), and photosynthetically active radiation (PAR) were selected. Only reliable data that passed the official quality control flags were retained; reliable data was defined as data with an official quality standard of 1. The PAR profile depth ranged from 0 to 200 meters. The surface photosynthetically active radiation was extracted from rice, and the chlorophyll concentration profile was set at a depth of 0 to 300 meters. Rice extract surface chlorophyll concentration, set Extract sea surface temperature.

[0100] The deep learning model structure for dual-branch joint inversion in this invention is as follows: Figure 1As shown, the input vector includes Surf_CHL (surf chlorophyll concentration in the surface layer), Surf_PAR (surf photosynthetically active radiation in the surface layer), SST (sea surface temperature), Lon & Lat (coordinates), and Date & Time (time). The input vector is fed into the first and second branches respectively. The first branch uses a multilayer perceptron structure, including a feature extraction module, an output layer, and a dimension expansion layer. The second branch uses a hybrid structure of multilayer perceptron and bidirectional LSTM, including a feature extraction module, a feature fusion layer, a sequence modeling module, and an output layer. The feature extraction module of the first branch includes three fully connected layers: the first fully connected layer is a Linear (10→64) layer with Tanh activation function; the second fully connected layer is a Linear (64→1) layer... The second branch's feature extraction module consists of three fully connected layers: the first is a Linear (10→128) layer with Tanh activation, the second is a Linear (128→128) layer with Tanh activation, the third is a Linear (128→301) layer with Tanh activation, and the feature fusion layer is a concat layer. The output of the second branch is then reshaped. Reshape the output of the first branch as follows: Then, concatenate the last dimension to form the input sequence. The detach operation is used to decouple the gradient of the first branch's output, preventing the gradient of the second branch from backpropagating to the first branch. The sequence modeling module includes a forward LSTM layer and a backward LSTM layer. Each forward and backward LSTM outputs 32-dimensional features, which are concatenated to obtain 64 dimensions. The LSTM is used to capture depth dependencies and contextual information in the vertical profile. The output layer is a Linear (64→1) layer. Each depth point is predicted independently, and the final output... The dimension tensor corresponds to the predicted chlorophyll concentration profile from 0 to 300 meters.

[0101] The training strategy employs an end-to-end joint training approach, achieving collaborative optimization of the two branches through a separate loss function and gradient decoupling mechanism. The mean squared error (MSE) is used as the loss function for both branches, with the true values ​​as labels and the MSE calculated in conjunction with the predicted values. A gradient backpropagation strategy is employed, using a step-by-step backpropagation mechanism. The first step calculates... Call `backward(retain_graph=True)` for backpropagation, update the training parameters for the second branch, and calculate the second step. The `backward()` function is called to propagate back and update the training parameters of the first branch. The `detach()` operation ensures that the gradient of the second branch is not propagated back to the first branch, achieving unidirectional information transfer. The Adam optimizer is used with an initial learning rate of 1×10⁻³, leveraging its adaptive learning rate characteristic to achieve fast convergence.

[0102] To evaluate the model's generalization ability under independent data source conditions, a test set based on MODIS satellite observation data (MODIS test set) was further constructed. This test set is not used for model training; it is only used to verify the generalization performance of the trained model on satellite data. The preprocessing procedures for all MODIS input variables are completely consistent with those for the BGC-Argo data, including variable selection, linear normalization methods, and encoding methods for spatiotemporal features, to ensure consistency in the feature space across different data sources. The model output corresponds to BGC-Argo buoy chlorophyll vertical profile data that is strictly matched to MODIS observations in time and space. It is important to emphasize that this part of the BGC-Argo profile is only used for independent validation on the MODIS test set and is completely independent of the datasets used for model training and testing. Root mean square error was used. ,average Coefficient of determination and correlation coefficient Statistical indicators are used to quantitatively evaluate predictive performance.

[0103] The chlorophyll concentration profile is shown in Table 1:

[0104] Table 1. Evaluation Results of Chlorophyll Concentration Profile

[0105] ;

[0106] After training on the BGC-Argo dataset, the model still achieves prediction performance on the MODIS test set that is very close to that of the BGC-Argo test set. (The two datasets are mentioned separately.) The correlation coefficients ρ for the doses of 0.10 mg·m⁻³ and 0.12 mg·m⁻³ remained above 0.84, and the coefficient of determination was... The value was 0.71 in both test sets. This result indicates that the model trained based on BGC-Argo observation data can be effectively applied to MODIS satellite product inputs, maintaining stable inversion capabilities under different data sources and observation systems, demonstrating the model's good transferability and generalization performance.

[0107] The results of the photosynthetically active radiation profile assessment are shown in Table 2:

[0108] Table 2. Evaluation Results of Photosynthetically Active Radiation Profile

[0109] ;

[0110] Table 2 presents the PAR profile prediction evaluation results of the model on the BGC-Argo and MODIS test sets. Overall, the PAR prediction performance is significantly better than the chlorophyll (CHL) profile inversion results, exhibiting higher prediction accuracy in both test sets. Specifically, the correlation coefficient of the PAR profile... All reached 0.97, with a coefficient of determination. The values ​​of 0.94 and 0.95 respectively indicate that the model can accurately characterize the vertical structure of PAR as it varies with depth. Meanwhile, and The numerical values ​​on the BGC-Argo and MODIS test sets are close, showing stable prediction accuracy. The above results verify the robustness of the model in the PAR profile inversion task and also provide support for its promotion and application in satellite observation data.

[0111] Figure 2 This is one of the predicted results of the vertical distribution of chlorophyll concentration using the model of this invention on the BGC-Argo test sample. It can be seen that the CHL profile predicted by the model generally maintains good consistency with the measured Argo profile, and the peak depth and peak amplitude have a high degree of agreement. This sample corresponds to... The RMSE was 0.056 (mg / m³). 3 This indicates that the model has a relatively robust ability to reconstruct profiles on BGC-Argo data.

[0112] Figure 3 This is a comparative result of the CHL profile for another BGC-Argo test sample. The predicted curve shows a high degree of agreement with the observed results in the upper water layer, and the overall trend is consistent. The coefficient of determination for this sample reaches [value missing]. The RMSE was 0.059 (mg / m³). 3 ).

[0113] Figure 4 This paper demonstrates the predicted vertical distribution of photosynthetically active radiation (PAR) using the model of this invention on the BGC-Argo test sample. The PAR profile predicted by the model is highly consistent with the measured results from Argo over the entire depth range of 0–200 m, and the coefficient of determination for this sample reaches [value missing]. , RMSE is 10.055(microeinsteinm -2 s -1 This indicates that the model's ability to reconstruct the PAR profile is very stable under Argo data-driven conditions.

[0114] Figure 5This is a comparative result of the PAR profile from another BGC-Argo test sample. The model can also reproduce the vertical structure of PAR decreasing from the surface to the depth well, and maintain consistency with the measured values ​​across different depth ranges. Compared to Figure 4 The sample exhibits a slight bias in the shallow to mid-depth layers (approximately 30–60 m), but the overall trend and decay rate remain consistent with Argo observations. This sample corresponds to... , RMSE is 56.499(microeinsteinm -2 s -1 ).

[0115] Figure 6 This demonstrates the prediction results for the CHL vertical profile using MODIS remote sensing products as input. A comparison with Argo's measured profiles shows high consistency between the prediction results and the mid-water layer. This sample... The concentration reached 0.877, and the RMSE was 0.035 (mg / m³). 3 The overall error level is low.

[0116] Figure 7 Another MODIS test sample's CHL profile prediction results are presented. Although there are some deviations in the surface and deep layers, the predicted profile is consistent with Argo observations in terms of overall trend and key structures. This sample's... The RMSE was 0.034 (mg / m³). 3 This indicates that the model still possesses stable and reliable profile inversion performance even when relying solely on remote sensing input.

[0117] Figure 8 This paper presents the predicted vertical profile of PAR (Paracrylamide) using MODIS remote sensing products as input. Although MODIS data only provides sea surface information, the model is still able to reasonably infer the vertical distribution characteristics of the light field within the water body and successfully reconstruct the overall morphology of PAR decaying exponentially with depth. The predicted results generally maintain good agreement with the Argo measured profile. This sample... , RMSE is 57.956(microeinsteinm -2 s -1 This indicates that the model still has strong stability and applicability in cross-data source application scenarios.

[0118] Figure 9 Another MODIS sample's PAR profile prediction results are presented. The model effectively captures the typical vertical structural characteristics of PAR, which decays rapidly at the surface and tends to flatten out at deeper layers, and maintains consistency with the overall trend of the Argo measured profile. The coefficient of determination for this sample reaches [value missing]. , RMSE is 38.160(microeinsteinm-2 s -1 This further verifies the feasibility of the model inverting the PAR profile under conditions that rely solely on remote sensing input.

[0119] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A sea surface-driven method for reconstructing the vertical profile of marine chlorophyll and light field, characterized in that, include: S1. Set filtering conditions to filter Argo profiles. Argo profiles include chlorophyll concentration profiles, ocean temperature profiles, and photosynthetically active radiation profiles. Standardize the filtered Argo profiles and perform physical fitting to generate surface photosynthetically active radiation and surface chlorophyll concentration. Convert the latitude and longitude of the observation points to three-dimensional unit spherical coordinates. Extract the date and time information of the observation points within the year to construct a periodic code. Extract the sea surface temperature of the Argo profile. Linearly normalize the sea surface temperature, surface chlorophyll concentration, and surface photosynthetically active radiation to construct the input vector. S2. Construct a deep learning model for dual-branch joint inversion, train the neural network, take the input vector as input, and take the real photosynthetically active radiation profile and the real chlorophyll concentration profile as real labels. Constrain the deep learning model for dual-branch joint inversion through a segregated loss function, update the dual-branch training parameters through a step-by-step backpropagation mechanism, set an iteration number threshold, and output the trained dual-branch joint inversion deep learning model when the iteration number reaches the iteration number threshold. After the Argo profile to be predicted is processed in step S1, it is input into the trained dual-branch joint inversion deep learning model, and the final chlorophyll concentration profile prediction result is output. S2 includes constructing a dual-branch joint inversion deep learning model ZS-Net. ZS-Net consists of two branches: the first branch adopts a multilayer perceptron structure, including a feature extraction module, an output layer, and a dimension expansion layer; the second branch adopts a hybrid structure of multilayer perceptron and bidirectional LSTM, including a feature extraction module, a feature fusion layer, a sequence modeling module, and an output layer. The feature extraction module of the first branch includes three fully connected layers. Each fully connected layer includes a Linear layer and a Tanh activation function. The output layer is a Linear layer, and the dimension expansion layer is a Linear layer. The feature extraction module of the second branch includes three fully connected layers. Each fully connected layer includes a Linear layer and a Tanh activation function. The feature fusion layer is a concat layer. The sequence modeling module includes a forward LSTM layer and a backward LSTM layer. The output layer is a Linear layer.

2. The sea surface-driven method for reconstructing the vertical profile of marine chlorophyll and light field according to claim 1, characterized in that, S1 includes, S1.1, setting the photosynthetically effective radiation depth threshold. , obtain Photosynthetically active radiation to the surface Calculate the effective photosynthetic radiation of the surface layer using a quadratic function. : ； ； In the formula, , , The coefficients are quadratic polynomials. It is a natural exponential function; S1 includes S1.2, and the depth of the photosynthetically active radiation profile is... ,Will Linear interpolation to ,get And perform a logarithmic transformation: ； In the formula, Photosynthetically active radiation label; S1 includes, S1.3, performing point smoothing on the chlorophyll concentration profile, assuming the chlorophyll concentration profile depth is... Set chlorophyll concentration depth threshold , obtain chlorophyll concentration to the surface The chlorophyll concentration profile was characterized using a Gaussian parametric model: ； In the formula, These are chlorophyll concentration profile parameters. , This represents the background concentration of the chlorophyll concentration profile. The thickness of the subsurface chlorophyll layer is the maximum value. This represents the depth corresponding to the maximum chlorophyll concentration. The parameter is used to control the amplitude of the peak term in the vertical distribution of chlorophyll. Indicates depth as The unit is meters. For depth indexing, ; S1 includes, S1.4, and... Perform initialization, including Initialize to the minimum value of the chlorophyll concentration profile. Initialized as the difference between the maximum and minimum values ​​of the chlorophyll concentration profile. Set a fixed initial value. The depth corresponding to the maximum chlorophyll concentration; Use the first linear least squares method to initialize the... Estimate the optimal chlorophyll concentration profile parameters to obtain the optimal parameters. : ； In the formula, for Observed chlorophyll concentration at [location]; S1 includes, S1.5, from 0 to Depth sequence of rules constructed within the range , For chlorophyll depth index, ,Will Substitution ,calculate chlorophyll concentration value, Indicates depth as The unit is meters.

3. The sea surface-driven method for reconstructing the vertical profile of marine chlorophyll and light field according to claim 2, characterized in that, S1 includes S1.6, converting the latitude and longitude of Argo profile observation points into three-dimensional unit spherical coordinates: ； ； ； In the formula, The latitude of the observation point. The longitude of the observation point. The x-coordinate of the observation point. The y-coordinate of the observation point. The z-axis coordinates of the observation point; S1 includes S1.7, extracting the date and time information within the year from the observation time of the observation point, and using sine and cosine functions for periodic encoding: ； ； ； ； In the formula, This refers to the number of days in the year. For a moment, The sine function is used to periodically encode the dates within the year. The periodic encoding of dates within the year using the cosine function. The sinusoidal function is periodically encoded to represent the time information. The periodic encoding of time information is a cosine function.

4. The sea surface-driven method for reconstructing the vertical profile of marine chlorophyll and light field according to claim 3, characterized in that, S1 includes, S1.8, setting the sea surface temperature depth threshold. , obtain Sea surface temperature The linear normalization formula is constructed as follows: ； In the formula, These are the original values ​​of the corresponding variables. The minimum value of the corresponding data. The maximum value of the corresponding data. The normalized corresponding variables include... , and .

5. The sea surface-driven method for reconstructing the vertical profile of marine chlorophyll and light field according to claim 4, characterized in that, S1 includes, S1.9, using the linear normalization formula to... , and Linear normalization was performed to obtain the normalized sea surface temperature. Normalized sea surface chlorophyll concentration and normalized sea surface photosynthetically active radiation Construct the input vector : 。 6. The sea surface-driven method for reconstructing the vertical profile of marine chlorophyll and light field according to claim 5, characterized in that, The output of the first branch is input into the feature fusion layer of the second branch, and then concatenated in the last dimension by the concat layer to form a sequence input. , For batches, The depth of the chlorophyll profile is the number of layers; The input sequence modeling module outputs 32-dimensional features from the forward LSTM layer and the backward LSTM layer, which are then concatenated to obtain 64-dimensional features. The results of the sequence modeling module are input into the output layer of the second branch to output the chlorophyll concentration profile prediction results.

7. The sea surface-driven method for reconstructing the vertical profile of marine chlorophyll and light field according to claim 6, characterized in that, Using true photosynthetically active radiation and true chlorophyll concentration as true values, a segregated loss function is constructed, which includes a first branch loss. Second branch loss , This represents the mean square error of photosynthetically active radiation. This represents the mean square error of chlorophyll concentration.

8. The sea surface-driven method for reconstructing the vertical profile of marine chlorophyll and light field according to claim 7, characterized in that, A step-by-step backpropagation mechanism is adopted. Update the training parameters for the first branch. Update the training parameters for the second branch; the second branch uses a detach operation to prevent... Backpropagation to the first branch.

Images