Inversion method of subsurface chlorophyll a concentration based on 1D-DSCAM

Abstract

The present invention relates to the field of chlorophyll a concentration prediction, and discloses a subsurface chlorophyll a concentration inversion method based on 1D-DSCAM, comprising the following steps: obtaining satellite remote sensing sea surface temperature and sea surface chlorophyll a concentration data and subsurface chlorophyll a concentration, subsurface temperature and salinity profile data of biogeochemical buoys, and preprocessing, constructing a data set, and inputting the training set into a 1D-DSCAM-GRU deep learning model for model training; the model comprises an Embedding module, a 1D-DSCAM module and a multi-layer GRU module; after preprocessing the sea surface temperature and sea surface chlorophyll a concentration data obtained by satellite remote sensing of the sea area to be predicted and the subsurface temperature and salinity profile data obtained by the biogeochemical buoys, the data are input into a qualified model for subsurface chlorophyll a concentration prediction. The method disclosed by the present invention can improve the accuracy of prediction.

Description

Technical Field

[0001] The invention relates to the field of chlorophyll a concentration prediction, and in particular to a subsurface chlorophyll a concentration inversion method based on 1D-DSCAM. Background Art

[0002] At present, the inversion of subsurface chlorophyll a concentration mainly relies on two data sources: remote sensing data and in situ data. Remote sensing data can provide a wide range of sea surface environmental elements, such as sea surface temperature (SST) and sea surface chlorophyll a concentration, but its limitation is that it cannot directly provide information below the surface; in situ data are obtained through coastal monitoring projects, limited-term marine surveys, and platforms such as biogeochemical buoys (Biogeochemical Argo, BGC-Argo). Although these data provide a detailed local description of each layer of the water column, their wide coverage in time and space is relatively limited.

[0003] To solve this problem, researchers combined these two types of data and used machine learning methods, especially artificial neural networks, to invert subsurface chlorophyll a concentrations. This method uses satellite-observed surface variables as input, matches them with in-situ measured subsurface chlorophyll a concentration profiles, and trains the neural network model to learn the mapping relationship of inferring subsurface characteristics from surface observations. This not only fully utilizes the advantages of the wide coverage of remote sensing data, but also combines the high-precision characteristics of in-situ data, effectively improving the accuracy of predictions.

[0004] At present, the technology of inverting chlorophyll a vertical profile using satellite data has been widely used. However, the research on inverting chlorophyll a vertical profile based on temperature and salinity profile data obtained by BGC-Argo floats is relatively limited. Even in existing applications, it is rare to see in-depth feature extraction and analysis of temperature and salinity profile data. In addition, the existing methods for inverting chlorophyll a vertical profile generally have the problem of insufficient training for small-scale phenomena, that is, most data sets are concentrated in the range of 0-0.5 For the range above 0.5 In the case of high concentration, it is difficult to obtain sufficient training effect. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides an inversion method for subsurface chlorophyll a concentration based on 1D-DSCAM to achieve the purpose of improving prediction accuracy.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] A method for inverting subsurface chlorophyll a concentration based on 1D-DSCAM comprises the following steps:

[0008] Step 1: obtain the sea surface temperature and sea surface chlorophyll a concentration data from satellite remote sensing and the subsurface chlorophyll a concentration, subsurface temperature and salinity profile data from biogeochemical buoys, and preprocess them. The preprocessed longitude, latitude, year, sampling date, sea surface temperature, sea surface chlorophyll a concentration, subsurface temperature and salinity profile data are integrated into the input feature matrix of the model. The subsurface chlorophyll a concentration profile data is used as the target output of the model. A data set is constructed and divided into a training set and a test set.

[0009] Step 2: Input the training set into the 1D-DSCAM-GRU deep learning model for model training, and use the data of the test set to evaluate the model accuracy;

[0010] The 1D-DSCAM-GRU deep learning model includes an Embedding module, a 1D-DSCAM module and a multi-layer GRU module;

[0011] The longitude, latitude, year, and sampling date in the constructed input feature matrix are input into the Embedding module for dimension upgrading; the profile data of temperature and salinity are stacked along the channel dimension to form the initial dual-channel input, which is input into the 1D-DSCAM module for processing. The data output from the 1D-DSCAM module and the sea surface temperature, sea surface chlorophyll a concentration, and data processed by the Embedding module are connected along the feature dimension and input into the multi-layer GRU module for prediction;

[0012] The 1D-DSCAM module includes a channel attention module and a spatial attention module. The input data first passes through the channel attention module, and then passes through the average pooling layer and the maximum pooling layer respectively, and then enters the multilayer perceptron 1 and the multilayer perceptron 2 respectively. Then, the outputs of the two multilayer perceptrons are added and passed through the sigmoid function to generate a channel attention weight between 0 and 1; the weighted signal is then passed through the spatial attention module, and the average value and the maximum value of the channel after passing through the channel attention module are respectively calculated. After the average value and the maximum value features are spliced, they are passed through a one-dimensional convolution layer, and the output is then passed through the sigmoid function to generate the spatial attention weight; the final output is the result of multiplying the input feature by the channel attention weight and the spatial attention weight;

[0013] Step 3: After preprocessing the sea surface temperature and sea surface chlorophyll a concentration data obtained by satellite remote sensing of the sea area to be predicted, as well as the subsurface temperature and salinity profile data obtained by biogeochemical buoys, they are input into the qualified model for subsurface chlorophyll a concentration prediction.

[0014] In the above scheme, the mixed loss function used in model training is as follows:

[0015] ;

[0016] in, = 0.5, used to balance the contribution of the two loss functions, N is the total number of all samples, is the Smooth L1 loss, defined as:

[0017] ;

[0018] in, = 0.5, is the predicted value, is the true value; is the weight of each sample The normalized result is Dynamic calculation based on the distribution of samples in different intervals.

[0019] In the above scheme, The calculation is as follows:

[0020] (1) Define the true value interval: divide the true value of subsurface chlorophyll a concentration into 0-0.5 , 0.5-1.0 , 1-1.5 Three intervals, and will be greater than 1.5 The concentration values ​​are uniformly classified as 1-1.5 in the category of;

[0021] (2) Calculate the number of samples in each interval ;

[0022] (3) Update the cumulative sum , using momentum smoothing mechanism to update the cumulative count of each interval:

[0023] ;

[0024] in, The role of is to help dynamically adjust the weight of samples in each interval, and the momentum parameter momentum=0.2;

[0025] (4) Calculate the sample weight in each interval , and normalize it:

[0026] ;

[0027] ;

[0028] in, is the total number of samples, is the sum of the number of samples in all intervals.

[0029] In the above scheme, the preprocessing in step 1 includes interpolation processing, integration matching, date conversion and normalization processing.

[0030] In a further technical solution, the interpolation processing is: the subsurface chlorophyll a concentration, subsurface temperature and salinity profile data are processed using a one-dimensional linear interpolation method. Taking the linear interpolation of the subsurface chlorophyll a concentration as an example, the specific method is as follows:

[0031] ;

[0032] in, , and , are the depth and subsurface chlorophyll a concentration of two known data points, is the depth point that needs to be interpolated, is the interpolated subsurface chlorophyll a concentration value.

[0033] In a further technical solution, the integration and matching is as follows: the sea surface temperature data and the sea surface chlorophyll a concentration data are integrated with the subsurface chlorophyll a concentration, subsurface temperature and salinity profile data using the nearest neighbor matching method, specifically as follows: the longitude and latitude coordinates extracted from the subsurface chlorophyll a concentration profile data are regarded as the matching center point, and the data of the observation point closest to the center point are screened out from the sea surface temperature and sea surface chlorophyll a concentration data obtained by satellite remote sensing as the data for the subsequent input model.

[0034] In a further technical solution, the date conversion is as follows: the sampling date extracted from the subsurface chlorophyll a concentration profile data is converted into a year and the position of the corresponding date in a year expressed in radians. :

[0035] ;

[0036] in, Refers to the day of the year, with a value range of 1-365, and 366 if it is a leap year.

[0037] In a further technical solution, the normalization process is as follows: the interpolated subsurface chlorophyll a concentration profile data is normalized, and the normalization formula is:

[0038] ;

[0039] Where Z represents the normalized subsurface chlorophyll a concentration value; is the actual observed value of subsurface chlorophyll a concentration after interpolation; and represent the mean and standard deviation of the observations, respectively.

[0040] In the above scheme, the multilayer perceptron 1 and the multilayer perceptron 2 are both composed of two one-dimensional convolutional layers connected in series and a RELU activation function. The multilayer perceptron 1 has been scaled by 8 times, and the multilayer perceptron 2 has been scaled by 16 times. The scaling is achieved through two 1×1 one-dimensional convolutional layers, that is, the first one-dimensional convolutional layer of the multilayer perceptron 1 reduces the input 128 channels to 16 in the channel dimension, and the second convolutional layer restores 128 channels, while the first convolutional layer of the multilayer perceptron 2 reduces the input 128 channels to 8 in the channel dimension, and the second convolutional layer restores 128 channels.

[0041] In the above scheme, the multi-layer GRU module has a four-layer structure. At each time step, the model receives an input vector and processes it through a four-layer GRU network. Each layer of the network contains 300 hidden state units, thereby achieving effective learning and processing of sequence data.

[0042] The GRU network has two gating mechanisms: update gate and reset gate;

[0043] The update gate determines how much information of the hidden state at the previous moment needs to be retained to the current moment. The calculation formula of the update gate is as follows:

[0044] ;

[0045] in, is the output of the update gate; It is the Sigmoid activation function, and the output value is between 0 and 1. is the weight matrix of the update gate; is the weight matrix from hidden state to update gate; is the bias term of the update gate; is the hidden state at the previous moment; is the input at the current moment;

[0046] The reset gate determines how much information of the hidden state at the previous moment needs to be forgotten. The calculation formula of the reset gate is as follows:

[0047] ;

[0048] in, is the output of the reset gate; is the weight matrix of the reset gate; is the weight matrix from hidden state to reset gate; is the bias term for the reset gate;

[0049] Candidate hidden states It is calculated based on the current input and the hidden state of the previous moment. The calculation formula is as follows:

[0050] ;

[0051] in, It is a hyperbolic tangent activation function, and the output value is between -1 and 1; is the weight matrix of the candidate hidden states; is the weight matrix from hidden state to candidate hidden state; is the bias term of the candidate hidden state;

[0052] The hidden state at the current moment It is a weighted combination of the previous hidden state and the candidate hidden state. The weight is controlled by the update gate. The calculation formula is as follows:

[0053] .

[0054] Through the above technical solution, the inversion method of subsurface chlorophyll a concentration based on 1D-DSCAM provided by the present invention has the following beneficial effects:

[0055] The present invention uses satellite remote sensing data and temperature and salinity vertical profile data obtained from BGC-Argo to predict the subsurface chlorophyll a concentration. First, the temperature and salinity profiles are deeply extracted. Existing feature extraction methods include CNN, Transformer, etc. The present invention innovatively uses a one-dimensional dual-scale channel attention module (1D Dual-Scale Channel Attention Module, 1D-DSCAM) as a feature extraction method. 1D-DSCAM is a variant of CBAM (Convolutional Block Attention Module), which is optimized for feature extraction of temperature and salinity vertical profiles, and can improve the accuracy and effect of feature extraction. In addition, a newly designed loss function, MixedTDWSmoothL1Loss, is introduced in the model training stage. This loss function combines TDWLoss (Target Distributed Weighted Loss) and Smooth L1 Loss. This loss function can enhance the importance of small-scale phenomena during training, reduce overfitting of common target values, make model learning more balanced, and adapt to changes in data distribution, stabilize the training process, and improve the generalization ability and robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art are briefly introduced below.

[0057] Figure 1 A schematic flow chart of a method for inverting subsurface chlorophyll a concentration based on 1D-DSCAM disclosed in an embodiment of the present invention;

[0058] Figure 2 The overall structure of the 1D-DSCAM-GRU deep learning model disclosed in the embodiment of the present invention;

[0059] Figure 3 It is a schematic diagram of CBAM;

[0060] Figure 4 A schematic diagram of a 1D-DSCAM module disclosed in an embodiment of the present invention;

[0061] Figure 5 This is a diagram of the GRU network structure disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0063] The present invention provides a method for inverting the subsurface chlorophyll a concentration based on 1D-DSCAM, such as Figure 1 As shown, the following steps are included:

[0064] Step 1: Obtain the satellite remote sensing sea surface temperature and sea surface chlorophyll a concentration data and the subsurface chlorophyll a concentration, subsurface temperature and salinity profile data of the biogeochemical buoy (BGC-Argo), and preprocess them. Integrate the preprocessed longitude, latitude, year, sampling date, sea surface temperature, sea surface chlorophyll a concentration, subsurface temperature and salinity profile data as the input feature matrix of the model. The subsurface chlorophyll a concentration profile data is used as the target output of the model. Construct a data set and divide it into a training set and a test set.

[0065] In this embodiment, the time range of the acquired data is from January 1, 2018 to March 1, 2022; the spatial range is the North Indian Ocean area of ​​0°N-30°N, 40°E-100°E.

[0066] Preprocessing includes interpolation, integration matching, date conversion and normalization.

[0067] (1) Interpolation processing

[0068] The subsurface chlorophyll a concentration, subsurface temperature and salinity profile data are processed using a one-dimensional linear interpolation method. Taking the linear interpolation of subsurface chlorophyll a concentration as an example, the specific method is as follows:

[0069] ;

[0070] in, , and , are the depth and subsurface chlorophyll a concentration of two known data points, is the depth point that needs to be interpolated, is the interpolated subsurface chlorophyll a concentration value.

[0071] (2) Integration and matching

[0072] The nearest neighbor matching method is used to effectively integrate the sea surface temperature data and sea surface chlorophyll a concentration data with the subsurface chlorophyll a concentration, subsurface temperature and salinity profile data, as follows: the longitude and latitude coordinates extracted from the subsurface chlorophyll a concentration profile data are regarded as the matching center point, and the data of the observation point closest to the center point are screened out from the sea surface temperature and sea surface chlorophyll a concentration data obtained by satellite remote sensing as the data for subsequent input into the model.

[0073] (3) Date conversion

[0074] The conversion of dates to radians allows for more efficient mathematical modeling and analysis of seasonal characteristics. The sampling dates extracted from the subsurface chlorophyll a concentration profile data were converted to years and the positions of the corresponding dates in radians in a year. :

[0075] ;

[0076] in, Refers to the day of the year, with a value range of 1-365, and 366 if it is a leap year.

[0077] (4) Normalization

[0078] The interpolated subsurface chlorophyll a concentration profile data is normalized, and the normalization formula is:

[0079] ;

[0080] Where Z represents the normalized subsurface chlorophyll a concentration value; is the actual observed value of subsurface chlorophyll a concentration after interpolation; and represent the mean and standard deviation of the observations, respectively.

[0081] (5) Building a dataset

[0082] After the above processing, the longitude, latitude, year, , sea surface temperature data, sea surface chlorophyll a concentration, subsurface temperature and salinity profile data are integrated into the model's input feature matrix, and the subsurface chlorophyll a concentration profile data is used as the model's target output. In the dataset construction phase, each row of data represents an independent observation sample, which contains all the above input features and their corresponding subsurface chlorophyll a concentration profile.

[0083] In order to evaluate the generalization ability of the model, the cross-validation method is used to divide the data set into a training set and a test set. Specifically, the data set is randomly split, with 80% of the data used as the training set for model training and parameter optimization; the remaining 20% ​​is used as the test set to evaluate the performance of the model on unseen data.

[0084] Step 2: Input the training set into the 1D-DSCAM-GRU deep learning model for model training, and use the data of the test set to evaluate the model accuracy.

[0085] (I) Building a 1D-DSCAM-GRU deep learning model

[0086] like Figure 2 As shown in the figure, the constructed 1D-DSCAM-GRU deep learning model includes an Embedding module, a 1D-DSCAM module and a multi-layer GRU module. The longitude, latitude, year and sampling date in the constructed input feature matrix are input into the Embedding module for dimension upgrading; the profile data of temperature and salinity are stacked along the channel dimension to form an initial dual-channel input. First, the number of channels of the data is expanded to 128 through convolution layer processing to enhance the model's ability to capture features, and then input into the 1D-DSCAM module to further improve the richness of feature expression. Finally, the convolution layer is used again to restore the number of output channels to the original two, ensuring the consistency of the data structure and the convenience of subsequent processing; the data output from the 1D-DSCAM module and the sea surface temperature, sea surface chlorophyll a concentration and the data processed by the Embedding module are connected together along the feature dimension and input into the multi-layer GRU module for prediction.

[0087] (1) Embedding module

[0088] For longitude, latitude, year, and The data is processed using the Embedding module, which can convert discrete, high-dimensional information such as time and space information into low-dimensional, dense continuous vectors to learn the potential relationship between the four variables. In addition, since the input of the Embedding module is usually an integer, for data with decimal points, the integer and decimal parts are separated, and the integer part is combined with the decimal part after passing through the Embedding module. This module is a prior art and will not be described in detail here.

[0089] (2) 1D-DSCAM module

[0090] The 1D-DSCAM module is improved based on the CBAM (Convolutional Block Attention Module). The CBAM is composed of Figure 3 As shown in the figure, it consists of two parts: the channel attention module and the spatial attention module. The channel attention module is input into the same multi-layer perceptron (MLP) after the average pooling layer and the maximum pooling layer. The MLP here consists of two one-dimensional convolutional layers (Conv1D) and a RELU activation function. The channel is reduced and restored through two series-connected one-dimensional convolutional layers (Conv1D) in the MLP to learn the dependencies between channels. The output is added and then passed through the sigmoid function to generate an attention weight between 0 and 1, which is used to weight each channel of the original input. The spatial attention module calculates the average and maximum values ​​of the channel respectively to capture the importance of different positions. After concatenating the average and maximum features, a one-dimensional convolutional layer (Conv1D) is used to learn the dependencies on the positions. The output also passes through the sigmoid function to generate the spatial attention weight.

[0091] Thus, in CBAM, the input first passes through the channel attention module to generate channel-level attention weights. The weighted signal then passes through the spatial attention module to generate spatial-level attention weights. The final output is the result of multiplying the input feature by the channel and spatial attention weights.

[0092] The 1D-DSCAM module constructed by the present invention is as follows Figure 4 As shown in FIG, similar to CBAM, it includes a channel attention module and a spatial attention module. After the input features are weighted by these two modules, the output is the result of the multiplication of the weights of the two modules. However, the present invention improves the channel attention module by inputting the outputs of the average pooling layer and the maximum pooling layer into two different MLPs respectively, so that more profile features can be obtained.

[0093] Specifically, the input data first passes through the channel attention module, and then passes through the average pooling layer and the maximum pooling layer, and then enters the multilayer perceptron 1 (MLP1) and multilayer perceptron 2 (MLP2). Then the outputs of the two multilayer perceptrons are added and passed through the sigmoid function to generate a channel attention weight between 0 and 1; the weighted signal then passes through the spatial attention module to calculate the average and maximum values ​​of the channels after passing through the channel attention module, and after concatenating the average and maximum value features, a one-dimensional convolutional layer (Conv1D) is used to learn the positional dependency, and the output passes through the sigmoid function to generate the spatial attention weight; the final output is the result of multiplying the input feature by the channel attention weight and the spatial attention weight.

[0094] Both Multilayer Perceptron 1 and Multilayer Perceptron 2 consist of two serially connected one-dimensional convolutional layers (Conv1D) and RELU activation functions. Multilayer Perceptron 1 is scaled by 8 times, and Multilayer Perceptron 2 is scaled by 16 times. The scaling is achieved through two 1×1 one-dimensional convolutional layers, that is, the first one-dimensional convolutional layer of Multilayer Perceptron 1 reduces the 128 channels of the input to 16 in the channel dimension, and the second convolutional layer restores 128 channels, while the first convolutional layer of Multilayer Perceptron 2 reduces the 128 channels of the input to 8 in the channel dimension, and the second convolutional layer restores 128 channels. The outputs of the two MLPs are added and passed through the sigmoid function to generate attention weights between 0 and 1, which are used to weight each channel of the original input. The spatial attention module is consistent with CBAM.

[0095] (3) Multi-layer GRU module

[0096] The present invention uses a gated recurrent unit (GRU) with a multi-layer structure as the core component of sequence modeling. GRU is an efficient and flexible recurrent neural network (RNN) variant that aims to solve long-term dependency problems while reducing computational costs and model complexity, and can capture rich temporal features when processing long sequence data. The multi-layer design enables the model to learn multi-level abstract features and better understand the complex patterns in the input sequence. The output of each layer serves as the input of the next layer. The first layer receives the temperature and salinity profiles after feature extraction, SST without any operation, sea surface chlorophyll a, and spatiotemporal data after the Embedding module. These data are spliced ​​into the input variables of the GRU layer according to the feature dimension. The output of the last layer of GRU is the data we need to predict.

[0097] The multi-layer GRU module has a four-layer structure. At each time step, the model receives the input vector and processes it through a four-layer GRU network. Each layer of the network contains 300 hidden state units, thereby achieving effective learning and processing of sequence data; Figure 5As shown, for each element of the input sequence, each layer of the GRU network calculates as follows:

[0098] The GRU network has two gating mechanisms: update gate and reset gate;

[0099] The update gate determines how much information of the hidden state at the previous moment needs to be retained to the current moment. The calculation formula of the update gate is as follows:

[0100] ;

[0101] in, is the output of the update gate; It is the Sigmoid activation function, and the output value is between 0 and 1. is the weight matrix of the update gate; is the weight matrix from hidden state to update gate; is the bias term of the update gate; is the hidden state at the previous moment; is the input at the current moment;

[0102] The reset gate determines how much information of the hidden state at the previous moment needs to be forgotten. The calculation formula of the reset gate is as follows:

[0103] ;

[0104] in, is the output of the reset gate; is the weight matrix of the reset gate; is the weight matrix from hidden state to reset gate; is the bias term for the reset gate;

[0105] Candidate hidden states It is calculated based on the current input and the hidden state of the previous moment. The calculation formula is as follows:

[0106] ;

[0107] in, It is a hyperbolic tangent activation function, and the output value is between -1 and 1; is the weight matrix of the candidate hidden states; is the weight matrix from hidden state to candidate hidden state; is the bias term of the candidate hidden state;

[0108] The hidden state at the current moment It is a weighted combination of the previous hidden state and the candidate hidden state. The weight is controlled by the update gate. The calculation formula is as follows:

[0109] .

[0110] 2. Model training

[0111] The preprocessed training set is input into the model to predict the chlorophyll a concentration profile, and the data set is divided into training set and test set in a ratio of 8:2. The batch size is set to 10, the training rounds are set to 300, the optimizer used is Adam, and the initial learning rate of the training is 0.0001. In order to prevent the model from overfitting during the training process, the present invention adopts an early stopping strategy, and continuously monitors the loss function value on the validation set during the training process as a reference indicator for early stopping. If the loss on the validation set does not improve significantly within 10 consecutive cycles, the training process will be terminated early.

[0112] In order to deal with the problem that the traditional loss function may cause the model to over-focus on common target values ​​and ignore small-scale phenomena, this paper designs an innovative loss function - MixedTDWSmoothL1Loss, which combines TDWLoss (Target Distributed Weighted Loss) and Smooth L1 Loss. The loss function is designed to solve the problem of unbalanced target value distribution, ensuring that small-scale phenomena can account for a larger proportion in loss calculation during model training, thereby improving the model's learning effect on these rare values ​​and ultimately improving overall performance.

[0113] Based on the given predicted value p and true value t, the mixed loss function is defined as:

[0114] ;

[0115] in, = 0.5, used to balance the contribution of the two loss functions, N is the total number of all samples, is the Smooth L1 loss, defined as:

[0116] ;

[0117] in, = 0.5, is the predicted value, is the true value; is the weight of each sample The normalized result is Dynamic calculation based on the distribution of samples in different intervals.

[0118] The calculation is as follows:

[0119] (1) Define the true value interval: divide the true value of subsurface chlorophyll a concentration into 0-0.5 , 0.5-1.0 , 1-1.5 Three intervals, and will be greater than 1.5 The concentration values ​​are uniformly classified as 1-1.5 in the category of;

[0120] (2) Calculate the number of samples in each interval ;

[0121] (3) Update the cumulative sum , using momentum smoothing mechanism to update the cumulative count of each interval:

[0122]

[0123] in, The role of is to help dynamically adjust the weight of samples in each interval, and the momentum parameter momentum=0.2;

[0124] (4) Calculate the sample weight in each interval , and normalize it:

[0125] ;

[0126] ;

[0127] in, is the total number of samples, is the sum of the number of samples in all intervals. Through normalization, the loss function can better balance the samples in each interval and improve the robustness and generalization ability of the model.

[0128] 3. Model evaluation

[0129] The present invention uses multiple parameters to evaluate the accuracy of its model and uses different models for comparison. The mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), coefficient of determination (R-squared, ) to evaluate the model accuracy. The following are the formulas for each indicator:

[0130] ;

[0131] ;

[0132] ;

[0133] in, is the true value, is the predicted value and n is the sample size.

[0134] The R² score represents the proportion of the total variation that is explained by the model. Its value ranges from 0 to 1, with values ​​closer to 1 indicating a better model fit.

[0135] ;

[0136] in, is the average of the true values.

[0137] The model accuracy is shown in Table 1.

[0138] Table 1 Model accuracy

[0139]

[0140] From Table 1, we can see that the 1D-DSCAM-GRU model has the highest accuracy and the smallest error, and its MAE is 0.054. , MSE is 0.011 , is 0.793, and RMSE is 0.106 In comparison, the CBAM-GRU model performs slightly worse, with a MAE of 0.054. , MSE is 0.014 , R² is 0.76, RMSE is 0.119 Compared with the 1D-DSCAM-GRU model, the MSE and RMSE of the CBAM-GRU model are slightly higher, and the R² is also slightly lower, indicating that the 1D-DSCAM module is better than the CBAM module in extracting the vertical profile characteristics of temperature and salinity, and the improvement of the model accuracy is more significant.

[0141] Compared with the GRU model that directly inputs satellite data and temperature-salinity profiles, the R² of the present invention is improved by 0.39. At the same time, compared with the GRU model that only inputs satellite data, the GRU model that directly inputs satellite data and temperature-salinity profiles shows higher errors and lower coefficients of determination. This shows that the vertical profile data of temperature and salinity can only significantly improve the prediction accuracy of the model after being processed by the feature extraction module. Directly inputting these data without feature extraction not only fails to improve the model performance, but may also lead to a decrease in the generalization ability of the model.

[0142] Step 3: After preprocessing the sea surface temperature and sea surface chlorophyll a concentration data obtained by satellite remote sensing of the sea area to be predicted, as well as the subsurface temperature and salinity profile data obtained by biogeochemical buoys, they are input into the qualified model for subsurface chlorophyll a concentration prediction.

[0143] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for inverting subsurface chlorophyll a concentration based on 1D-DSCAM, characterized in that: The steps include: Step 1: obtain the sea surface temperature and sea surface chlorophyll a concentration data from satellite remote sensing and the subsurface chlorophyll a concentration, subsurface temperature and salinity profile data from biogeochemical buoys, and preprocess them. The preprocessed longitude, latitude, year, sampling date, sea surface temperature, sea surface chlorophyll a concentration, subsurface temperature and salinity profile data are integrated into the input feature matrix of the model. The subsurface chlorophyll a concentration profile data is used as the target output of the model. A data set is constructed and divided into a training set and a test set. Step 2: Input the training set into the 1D-DSCAM-GRU deep learning model for model training, and use the data of the test set to evaluate the model accuracy; The 1D-DSCAM-GRU deep learning model includes an Embedding module, a 1D-DSCAM module and a multi-layer GRU module; The longitude, latitude, year, and sampling date in the constructed input feature matrix are input into the Embedding module for dimension upgrading; the profile data of temperature and salinity are stacked along the channel dimension to form the initial dual-channel input, which is input into the 1D-DSCAM module for processing. The data output from the 1D-DSCAM module and the sea surface temperature, sea surface chlorophyll a concentration, and data processed by the Embedding module are connected along the feature dimension and input into the multi-layer GRU module for prediction; The 1D-DSCAM module includes a channel attention module and a spatial attention module. The input data first passes through the channel attention module, and then passes through the average pooling layer and the maximum pooling layer, and then enters the multilayer perceptron 1 and the multilayer perceptron 2 respectively. Then the outputs of the two multilayer perceptrons are added and passed through the sigmoid function to generate a channel attention weight between 0 and 1; The weighted signal then passes through the spatial attention module to calculate the average and maximum values ​​of the channels after the channel attention module. The average and maximum value features are concatenated and passed through a one-dimensional convolution layer. The output then passes through the sigmoid function to generate the spatial attention weight. The final output is the result of multiplying the input feature by the channel attention weight and the spatial attention weight. Step 3: After preprocessing the sea surface temperature and sea surface chlorophyll a concentration data obtained by satellite remote sensing of the sea area to be predicted and the subsurface temperature and salinity profile data obtained by biogeochemical buoys, they are input into the qualified model for subsurface chlorophyll a concentration prediction.

2. The inversion method of subsurface chlorophyll a concentration based on 1D-DSCAM according to claim 1, characterized in that: The mixed loss function used in model training is as follows: ； in, = 0.5, which is used to balance the contribution of the two loss functions. N is the total number of samples. is the Smooth L1 loss, defined as: ； in, = 0.5, is the predicted value, is the true value; is the weight of each sample The normalized result is Dynamic calculation based on the distribution of samples in different intervals.

3. The inversion method of subsurface chlorophyll a concentration based on 1D-DSCAM according to claim 2, characterized in that: The calculation is as follows: (1) Define the true value interval: divide the true value of subsurface chlorophyll a concentration into 0-0.5 , 0.5-1.0 , 1-1.5 Three intervals, and will be greater than 1.5 The concentration values ​​are uniformly classified as 1-1.5 in the category of; (2) Calculate the number of samples in each interval ; (3) Update the cumulative sum , using momentum smoothing mechanism to update the cumulative count of each interval: ； in, The role of is to help dynamically adjust the weight of samples in each interval, and the momentum parameter momentum=0.2; (4) Calculate the sample weight in each interval , and normalize it: ； ； in, is the total number of samples, is the sum of the number of samples in all intervals.

4. The inversion method of subsurface chlorophyll a concentration based on 1D-DSCAM according to claim 1, characterized in that: The preprocessing in step 1 includes interpolation, integration matching, date conversion and normalization.

5. The inversion method of subsurface chlorophyll a concentration based on 1D-DSCAM according to claim 4, characterized in that: The interpolation process is: using a one-dimensional linear interpolation method to process the subsurface chlorophyll a concentration, subsurface temperature and salinity profile data. Taking the linear interpolation of the subsurface chlorophyll a concentration as an example, the specific method is as follows: ； in, , and , are the depth and subsurface chlorophyll a concentration of two known data points, is the depth point that needs to be interpolated, is the interpolated subsurface chlorophyll a concentration value.

6. The inversion method of subsurface chlorophyll a concentration based on 1D-DSCAM according to claim 4, characterized in that: The integration matching is as follows: the sea surface temperature data and the sea surface chlorophyll a concentration data are integrated with the subsurface chlorophyll a concentration, subsurface temperature and salinity profile data using the nearest neighbor matching method, specifically as follows: the longitude and latitude coordinates extracted from the subsurface chlorophyll a concentration profile data are regarded as the matching center point, and the data of the observation point closest to the center point are screened out from the sea surface temperature and sea surface chlorophyll a concentration data obtained by satellite remote sensing as the data for subsequent input into the model.

7. The inversion method of subsurface chlorophyll a concentration based on 1D-DSCAM according to claim 4, characterized in that: The date conversion is as follows: the sampling date extracted from the subsurface chlorophyll a concentration profile data is converted into a year and the position of the corresponding date in radians in the year : ； in, Refers to the day of the year, with a value range of 1-365, and 366 if it is a leap year.

8. The inversion method of subsurface chlorophyll a concentration based on 1D-DSCAM according to claim 4, characterized in that: The normalization process is as follows: normalize the interpolated subsurface chlorophyll a concentration profile data, and the normalization formula is: ； Where Z represents the normalized subsurface chlorophyll a concentration value; is the actual observed value of subsurface chlorophyll a concentration after interpolation; and represent the mean and standard deviation of the observations, respectively.

9. The inversion method of subsurface chlorophyll a concentration based on 1D-DSCAM according to claim 1, characterized in that: The multilayer perceptron 1 and the multilayer perceptron 2 are both composed of two one-dimensional convolutional layers connected in series and a RELU activation function. The multilayer perceptron 1 has been scaled by 8 times, and the multilayer perceptron 2 has been scaled by 16 times. The scaling is achieved through two 1×1 one-dimensional convolutional layers, that is, the first one-dimensional convolutional layer of the multilayer perceptron 1 reduces the 128 input channels to 16 in the channel dimension, and the second convolutional layer restores the 128 channels, while the first convolutional layer of the multilayer perceptron 2 reduces the 128 input channels to 8 in the channel dimension, and the second convolutional layer restores the 128 channels.

10. The inversion method of subsurface chlorophyll a concentration based on 1D-DSCAM according to claim 1, characterized in that: The multi-layer GRU module has a four-layer structure. At each time step, the model receives an input vector and processes it through a four-layer GRU network. Each layer of the network contains 300 hidden state units, thereby achieving effective learning and processing of sequence data. The GRU network has two gating mechanisms: update gate and reset gate; The update gate determines how much information of the hidden state at the previous moment needs to be retained to the current moment. The calculation formula of the update gate is as follows: ； in, is the output of the update gate; It is the Sigmoid activation function, and the output value is between 0 and 1. is the weight matrix of the update gate; is the weight matrix from hidden state to update gate; is the bias term of the update gate; is the hidden state at the previous moment; is the input at the current moment; The reset gate determines how much information of the hidden state at the previous moment needs to be forgotten. The calculation formula of the reset gate is as follows: ； in, is the output of the reset gate; is the weight matrix of the reset gate; is the weight matrix from hidden state to reset gate; is the bias term for the reset gate; Candidate hidden states It is calculated based on the current input and the hidden state of the previous moment. The calculation formula is as follows: ； in, It is the hyperbolic tangent activation function, and the output value is between -1 and 1; is the weight matrix of the candidate hidden states; is the weight matrix from hidden state to candidate hidden state; is the bias term of the candidate hidden state; The hidden state at the current moment It is a weighted combination of the previous hidden state and the candidate hidden state. The weight is controlled by the update gate. The calculation formula is as follows: 。

Images