Satellite remote sensing prediction and reconstruction method of underwater temperature field based on Transformer architecture

Abstract

The application relates to a kind of underwater temperature field satellite remote sensing prediction reconstruction methods based on Transformer architecture, a variety of sea surface parameter information (position information, time information, sea surface temperature, sea level anomaly) established by satellite remote sensing data are combined with known temperature profile information to realize the reconstruction of underwater temperature field using marine surface parameters, the known temperature information corresponding to the random 0-3 depth layer is added to the training data, and the accuracy of the reconstruction result is improved compared with the traditional method.The application does not need to understand and analyze the complex ocean dynamics process, but can directly obtain the reconstructed temperature field data by using the corresponding observation data, with small calculation amount, simple process implementation and accurate result.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of ocean physics, ocean engineering and underwater acoustic engineering, and relates to a method for predicting and reconstructing an underwater temperature field based on a Transformer architecture, which is suitable for predicting and reconstructing an underwater temperature field using satellite remote sensing data. BACKGROUND

[0002] Although a variety of methods for reconstructing an underwater temperature field have been used in engineering practice, such as empirical function regression method, self-organizing neural network method and deep evidence regression network method, there are still serious technical challenges in reconstructing an underwater temperature field in deep sea complex waters. The root cause is that the existing temperature profile reconstruction methods have certain defects when facing the strong nonlinear process of complex ocean environment, or the accuracy of reconstruction is not enough, and there is a lack of prediction and reconstruction of complete profiles, etc. The specific analysis is as follows:

[0003] (1) Single empirical orthogonal function method. This method represents the temperature profile by empirical orthogonal function and empirical orthogonal function coefficient, and establishes the regression relationship between the sea surface remote sensing parameters and the empirical orthogonal function coefficient. The reconstruction accuracy of the temperature profile is related to the ocean dynamic process, spatial resolution, time resolution, sea surface remote sensing parameter observation accuracy, sea surface remote sensing parameter combination, etc., among which the ocean dynamic process and the sea surface remote sensing parameter are the key. In the sea area where eddies frequently occur and strong fronts exist, there is usually no significant correlation between the sea surface remote sensing parameter and the empirical orthogonal function coefficient, which will cause significant error in the reconstruction of the temperature profile.

[0004] (2) Self-organizing neural network method. This method represents the temperature profile by empirical orthogonal function and empirical orthogonal function coefficient, and establishes the regression relationship between the sea surface remote sensing parameters and the empirical orthogonal function coefficient through self-organizing neural network. The reconstruction accuracy of the temperature profile is related to the ocean dynamic process, spatial resolution, time resolution, sea surface remote sensing parameter observation accuracy, sea surface remote sensing parameter combination, etc., among which the ocean dynamic process and the sea surface remote sensing parameter are the key. Although this method has good prediction results in the sea area where eddies frequently occur and strong fronts exist, there are still large prediction errors in some cases.

[0005] (3) Deep evidence regression network method. This method uses deep evidence regression network to reconstruct the underwater temperature field by combining the position information, time information and corresponding sea surface remote sensing parameters (such as sea surface temperature and sea surface height) of each profile in the study area. It realizes the reconstruction of temperature profile in different spatio-temporal scales, and gives the uncertainty estimation and confidence, confidence interval of the reconstructed profile. Although the prediction results of this method are consistent with the true results, there are still deviations between the prediction results and the true results of about ±2℃ or even more, and the prediction accuracy needs to be improved.

[0006] In short, the single empirical orthogonal function method, self-organizing neural network method, deep evidence regression network method and the like have relatively large errors when reconstructing the temperature profile in the active area of the ocean dynamic process, and there is still a gap for actual needs. Therefore, new principles and technical approaches must be sought to improve the reconstruction accuracy using field observation data. Underwater robots and other underwater unmanned systems or submarines can obtain temperature information at certain depth layers in real time. These real temperature information has a good effect on improving the reconstruction accuracy of the underwater temperature field and reducing errors. SUMMARY

[0007] TECHNICAL PROBLEM

[0008] In order to avoid the shortcomings of the prior art, the present application proposes an underwater temperature field satellite remote sensing prediction reconstruction method based on the Transformer architecture, which makes up for the shortcomings of the prior art in estimating and predicting results. The present application combines temperature information at some known depth layers to improve prediction accuracy and is suitable for predicting and reconstructing the underwater temperature field in the deep-sea multi-scale marine environment.

[0009] TECHNICAL SCHEME

[0010] An underwater temperature field satellite remote sensing prediction reconstruction method based on the Transformer architecture, characterized in that the steps are as follows:

[0011] Step 1: Establish a data set of the input feature network of the multi-dimensional information set, including the position information, time information and corresponding sea surface remote sensing parameter data of the reconstructed profile;

[0012] The sea surface remote sensing parameters include sea surface temperature data and sea level height anomaly data;

[0013] The position information includes longitude and latitude;

[0014] The time information is the month;

[0015] The position information data and the time information data are sine and cosine coded to convert them into data that can be understood and learned by the model as input data;

[0016] Step 2: Encapsulate the input data into a PyTorch data set, and stack the input features and temperature profiles of the data set into two-dimensional tensors;

[0017] Randomly select m known temperature layer temperature values, save the temperature values and their corresponding temperature layer indexes, process the data in the batch and mask the positions of the known temperatures;

[0018] Step 3, input embedding and position encoding: map the input features to a high-dimensional space through a linear layer, add position encoding to each position, even positions are encoded using the sine function, and odd positions are encoded using the cosine function; the value of the known temperature is mapped to a high-dimensional space after the temperature value is mapped to a high-dimensional space, and the position encoding and feature embedding are combined;

[0019] Step 4: load the Transformer model into the network by inputting the packaged PyTorch dataset in batches, use the unique multi-head attention mechanism of the Transformer model to complete the training of the model using historical data;

[0020] Step 5: input the position information, time information and corresponding sea surface remote sensing parameters such as sea surface temperature and sea level height anomaly of the reconstructed profile into the Transformer network, and obtain the temperature field reconstruction value and its corresponding MSE:

[0021]

[0022] where n represents the number of samples, y i represents the true value, is a prediction value; the original temperature profile and the reconstructed temperature profile are obtained by visualizing the test set data.

[0023] The m of the temperature values of the n known temperature layers is 1-3.

[0024] The encoding process of step 3 is as follows:

[0025]

[0026] where PE represents the position, used to map the position in the input sequence to the index of the position matrix, pos represents the position of the input data in the sequence, i represents the dimension index, ranging from 0 to d model represents the dimension of the model embedding.

[0027] The step 4 uses historical data to complete the training of the model, and the parameter settings during training are as follows: the loss function uses the mean square error loss function to calculate the difference between the predicted temperature of the model and the actual temperature; the optimizer uses AdamW, the learning rate is set to 10 -4 , and the weight decay is set to 10 -5 .

[0028] The learning rate scheduler uses ReduceLROnPlateau to automatically adjust the learning rate according to the loss of the test set.

[0029] The specific settings of the learning rate scheduler are as follows: when the test set loss does not decrease within 3 training rounds, the learning rate is reduced to half of the previous one; the entire model training round is set to 50 rounds with an early stopping mechanism. During the training process, the best loss of the test set up to the current training round is updated. If the best loss does not improve within 5 consecutive training rounds, the early stopping mechanism is triggered and training is stopped to prevent overfitting; in addition, the batch size is set to 16.

[0030] During the model training process, information of 1 to 3 known temperature layers is randomly added.

[0031] An electronic device, characterized in that it includes a processor and a memory, wherein the processor is used to implement the data migration steps of the underwater temperature field satellite remote sensing prediction and reconstruction method based on the Transformer architecture when executing the computer program stored in the memory.

[0032] A readable storage medium, characterized in that a computer program is stored on the readable storage medium, and when the computer program is executed by a processor, the steps of data migration of the underwater temperature field satellite remote sensing prediction and reconstruction method based on the Transformer architecture are implemented.

[0033] A computer program product, characterized in that it includes computer-executable instructions, which, when executed, are used to implement the underwater temperature field satellite remote sensing prediction and reconstruction method based on the Transformer architecture.

[0034] Beneficial effects

[0035] The present invention proposes a method for satellite remote sensing prediction and reconstruction of underwater temperature fields based on a Transformer architecture. It utilizes a variety of sea surface parameter information (position information, time information, sea surface temperature, and sea level height anomalies) established by satellite remote sensing data and previously known temperature profile information for joint inversion, realizing the reconstruction of the underwater temperature field using ocean surface parameters. By adding known temperature information corresponding to random 1-3 depth layers to the training data, the accuracy of the reconstruction results is improved compared with traditional methods.

[0036] This paper uses the Transformer architecture to establish a mapping network between sea surface parameters (position information, time information, sea surface temperature, and sea level height anomalies) and underwater temperature profiles based on a large amount of remote sensing data. The self-attention mechanism introduced by the Transformer architecture assigns different weights by calculating the similarity between each parameter in the input features and other parameters, thereby achieving refined processing of the input features and allowing the model to focus on the important information in the input features. Compared with traditional methods, it is more suitable for nonlinear mapping from sea surface parameters to underwater temperature profiles.

[0037] The application directly obtains reconstructed temperature field data by using corresponding observation data without understanding and analyzing complex ocean dynamics, has small calculation amount, simple process implementation, and relatively accurate result. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 : (a) shows the positions of measured temperature profiles near the Philippine Sea during 2001-2018 (longitude range: 123-135°E, latitude range: 15-25°N); (b) shows the temperature profiles used in constructing the Transformer network model in this sea area; (c) shows the positions of measured temperature profiles near the sea area east of Japan during 2001-2018 (longitude range: 145-165°E, latitude range: 30-45°N); (d) shows the temperature profiles used in constructing the Transformer network model in this sea area.

[0039] Figure 2 : Flow chart of the temperature field reconstruction framework based on the Transformer model: the left part of the flow chart is training the Transformer network by using training samples; the right part of the flow chart is predicting the temperature field by using known necessary information.

[0040] Figure 3 : (a), (b), (c), (d) show the comparison between the predicted profile and the observed profile when the number of known temperature points near the Philippine Sea is 0, 1, 2, and 3 respectively.

[0041] Figure 4 : (a), (b), (c), (d) show the comparison between the predicted profile and the observed profile when the number of known temperature points in the sea area east of Japan is 0, 1, 2, and 3 respectively. DETAILED DESCRIPTION

[0042] The application will be further described in combination with embodiments and drawings:

[0043] The application discloses a method for reconstructing an underwater temperature field based on a Transformer architecture, and has the characteristics that in a research sea area, a dataset of a Transformer feature input network of a multi-dimensional information set is established by using underwater temperature field data, combining position information, time information and corresponding sea surface remote sensing parameters such as sea surface temperature (SST) and sea level anomaly (SLA) data of each profile in the research sea area, and temperature data of each profile is taken as an ideal output result of the input dataset. After the training of the Transformer model is completed, the position information, time information and corresponding sea surface remote sensing parameters such as sea surface temperature and sea level anomaly information of a profile to be reconstructed can be inputted to inversely obtain a reconstructed temperature profile and mean squared error (MSE) information of the reconstructed temperature profile and real profile data. In order to improve the accuracy of the reconstruction result, 1-3 temperature information of a known depth layer is randomly added in the training process. The whole reconstruction process and principle can be divided into the following five steps.

[0044] Step one: a dataset of an input feature network of a multi-dimensional information set is established, including position information, time information and corresponding sea surface remote sensing parameter (sea surface temperature and sea level anomaly) data of a profile. The position information (here, longitude and latitude) and time information (here, month) are subjected to sine and cosine coding and are converted into data that can be understood and learned by a model, so as to capture the periodic characteristics in the data and enable the model to better understand the change rule of the data. The input features are standardized to accelerate the model training and improve the generalization ability.

[0045] Step two: the input data is encapsulated into a PyTorch dataset to enable batch loading in training and testing. The input features and temperature profiles are stacked into two-dimensional tensors, 1-3 temperature values of a known temperature layer are randomly selected, the temperature values and the corresponding temperature layer indexes are saved, the data in a batch is processed, and the positions of the known temperatures are subjected to mask processing.

[0046] Step three: input embedding and position coding. The input features are mapped to a high-dimensional space through a linear layer, and since the Transformer does not have implicit sequential information, position coding needs to be added to each position. In the Transformer architecture, even positions are coded by using a sine function, and odd positions are coded by using a cosine function. After the known temperature values are mapped to a high-dimensional space, the values are combined with the position coding and feature embedding. The coding process is as follows:

[0047]

[0048] Among them, PE represents position, which is used to map the position in the input sequence to the index of the position matrix, pos represents the position of the input data in the sequence, and i represents the dimension index, ranging from 0 to d model Represents the dimension of the model embedding.

[0049] Step 4: Based on the input features, a network model is established, and the network parameters are trained using historical data through the unique multi-head attention mechanism of the Transformer model.

[0050] Step 5: Input the Transformer network based on the position information, time information, and corresponding sea surface remote sensing parameters such as sea surface temperature and sea level height anomaly of the profile to be reconstructed to obtain the temperature field reconstruction value and its corresponding MSE, which can be expressed as:

[0051]

[0052] Among them, n represents the number of samples, y i represents the true value, Is the predicted value. The original temperature profile and the reconstructed temperature profile are obtained by visualizing the test set data.

[0053] The parameters in network training are set as follows: the loss function uses the mean square error loss function to calculate the difference between the model's predicted temperature and the actual temperature. The optimizer uses AdamW, and the learning rate is set to 10 -4 , weight decay is set to 10 -5 The learning rate scheduler used is ReduceLROnPlateau, which automatically adjusts the learning rate based on the loss of the test set. Specifically, when the loss of the test set does not decrease within three training rounds, the learning rate is reduced to half of the previous value. The entire training round of the model is set to 50 times and is accompanied by an early stopping mechanism. During the training process, the best loss of the test set up to the current training round is updated. When the best loss does not improve within five consecutive training rounds, the early stopping mechanism is triggered and training is stopped to prevent overfitting. In addition, the batch size is set to 16.

[0054] During the training process, the model will randomly add information about 1-3 known temperature layers. The larger the value, the better. The hope is to reconstruct the temperature field with information from fewer known temperature layers. The 0-3 here is to unify with the 0-3 of the test set, because we want to know whether the test set results are significantly different under different numbers of known temperature layers. Due to the random relationship, only a part of the data is not added to the known temperature layer during the training process. If the value of 0 is not appropriate here, changing it to 1-3 will indeed be better for the results.

[0055] Figure 1 In particular, (a) shows the positions of the measured temperature profiles in the Philippine Sea during 2001-2018, with a depth range of 0-1950 m; (b) shows the positions of the measured temperature profiles in the sea east of Japan during 2001-2018, with a depth range of 0-1950 m; (c) shows the temperature profiles in the Philippine Sea after removing some abnormal values for constructing the Transformer network; and (d) shows the temperature profiles in the sea east of Japan for constructing the Transformer network.

[0056] Figure 2 A framework flowchart based on the Transformer network is shown. The implementation process includes two processes:

[0057] (1) Sample training: The sea surface information (including position information, time information, sea surface temperature, and sea height anomaly) data are taken as inputs, and the corresponding profile data are taken as expected outputs for training the Transformer network. The total sample data used in the Philippine Sea is about 10146, and the total sample data used in the sea east of Japan is about 22114. 80% of the samples are randomly selected as training samples, and the remaining 20% of the samples are taken as test samples to verify the reconstruction performance of the algorithm.

[0058] (2) Reconstruction of the temperature field: The sea surface information (including position information, time information, sea surface temperature, and sea height anomaly) data are taken as inputs to obtain the reconstructed temperature field data. The reconstructed data are compared with the true profile temperature data, and the mean square error between the reconstructed data and the true data is calculated.

[0059] Figure 3 (a), (b), (c), and (d) show the comparison between the predicted profile and the observed profile when the number of known temperature points in the Philippine Sea is 0, 1, 2, and 3, respectively. The results show that the reconstructed temperature field has good consistency with the true temperature field data, and the mean square error value obtained by reconstruction in this sea area is about 0.2937.

[0060] Figure 4 (a), (b), (c), and (d) show the comparison between the predicted profile and the observed profile when the number of known temperature points in the sea east of Japan is 0, 1, 2, and 3, respectively. The results show that the reconstructed temperature field has good consistency with the true temperature field data, and the mean square error value obtained by reconstruction in this sea area is about 0.4062.

[0061] The present invention achieved good results in tests near the Philippine Sea and the sea east of Japan. Affected by the differences in the original data sets caused by the geographical location and environment of the sea area, the mean square error of the reconstruction results in the two sea areas was different, 0.2937 and 0.4062 respectively, and the average error temperature was within 0.6°C to 0.7°C, respectively, which is higher than the accuracy of traditional methods. The underwater temperature field reconstruction method based on the Transformer network does not require the understanding and analysis of complex ocean dynamic processes. It only needs to use the corresponding observation data to directly obtain the reconstructed temperature field data. The calculation amount is small, the process is simple to implement, and the results are relatively accurate. It is suitable for large-scale applications and is suitable for using satellite remote sensing observation data within a certain sea area to realize the reconstruction of the underwater temperature field in the area.

[0062] This method establishes a Transformer network with multi-dimensional information such as position information, time information, sea surface temperature, and sea level height anomaly. After the network training is completed, the reconstructed temperature field information under different depth conditions is obtained through the position information, time information and corresponding sea surface remote sensing parameters (sea surface temperature, sea level height anomaly) of the profile to be reconstructed.

[0063] Based on a large amount of remote sensing data, this paper establishes a mapping network between sea surface parameters (position information, time information, sea surface temperature, sea level height anomaly) and underwater temperature profiles, which can realize nonlinear mapping from sea surface parameters to underwater temperature profiles.

[0064] This method is based on big data training of the model. It does not require understanding and analysis of complex ocean dynamic processes. It only uses existing observation data and analyzes the correlation between different sea surface parameter data to reconstruct the underwater temperature field. It has a small amount of calculation, a simple process, and relatively accurate results. It is more suitable for using satellite remote sensing observation data in a certain sea area to reconstruct the underwater temperature field in the region.

Claims

1. A method for underwater temperature field satellite remote sensing prediction and reconstruction based on Transformer architecture, characterized by Here are the steps: Step 1: Establish a data set of the input feature network of the multi-dimensional information set, including the location information, time information of the reconstructed profile, and the corresponding sea surface remote sensing parameter data; The sea surface remote sensing parameters include sea surface temperature data and sea level height anomaly data; The location information includes longitude and latitude; The time information is a month; Perform sine and cosine encoding on the position information data and time information data, and convert them into data that can be understood and learned by the model as input data; Step 2: Encapsulate the input data into a PyTorch dataset, where the input features and temperature profiles of the dataset are stacked into a two-dimensional tensor. Randomly select m known temperature values ​​for each temperature layer, save the temperature value and its corresponding temperature layer index, process the data in the batch and mask the locations of known temperatures; Step 3: Input embedding and position encoding: The input features are mapped to a high-dimensional space through a linear layer. Position encoding is added to each position. Even positions are encoded using a sine function, and odd positions are encoded using a cosine function. After the known temperature value is mapped to the high-dimensional space, it is merged with the position encoding and feature embedding. Step 4: Input the encapsulated PyTorch dataset into the network in batches to load the Transformer model. Through the unique multi-head attention mechanism of the Transformer model, the model is trained using historical data. Step 5: The position information, time information, and corresponding sea surface remote sensing parameters such as sea surface temperature and sea level height anomalies of the profile to be reconstructed are input into the Transformer network to obtain the temperature field reconstruction value and its corresponding MSE: in, represents the number of samples, represents the true value, It represents the predicted value; the original temperature profile and the reconstructed temperature profile are obtained by visualizing the test set data.

2. The underwater temperature field satellite remote sensing prediction and reconstruction method based on the Transformer architecture according to claim 1 is characterized by: The temperature values ​​of the m known temperature layers are 1-3.

3. The underwater temperature field satellite remote sensing prediction and reconstruction method based on the Transformer architecture according to claim 1 is characterized by: The encoding process of step 3 is as follows: in, Represents the position, which is used to map the position in the input sequence to the index of the position matrix, Indicates the position of the input data in the sequence, Represents the dimension index, ranging from 0 to , Represents the dimension of the model embedding.

4. The underwater temperature field satellite remote sensing prediction and reconstruction method based on the Transformer architecture according to claim 1 is characterized by: Step 4 uses historical data to complete the training of the model. The parameters during training are set as follows: the loss function uses the mean square error loss function to calculate the difference between the model predicted temperature and the actual temperature; the optimizer uses AdamW, and the learning rate is set to 10 -4 , weight decay is set to 10 -5 .

5. The underwater temperature field satellite remote sensing prediction and reconstruction method based on the Transformer architecture according to claim 4 is characterized by: The learning rate scheduler uses ReduceLROnPlateau to automatically adjust the learning rate based on the loss of the test set.

6. The underwater temperature field satellite remote sensing prediction and reconstruction method based on the Transformer architecture according to claim 5 is characterized by: The specific settings of the learning rate scheduler are as follows: when the test set loss does not decrease within 3 training rounds, the learning rate is reduced to half of the previous one; the entire model training round is set to 50 rounds with an early stopping mechanism. During the training process, the best loss of the test set up to the current training round is updated. If the best loss does not improve within 5 consecutive training rounds, the early stopping mechanism is triggered and training is stopped to prevent overfitting; in addition, the batch size is set to 16.

7. An electronic device, characterized in that: It includes a processor and a memory, and the processor is used to implement the data migration steps of the underwater temperature field satellite remote sensing prediction and reconstruction method based on the Transformer architecture as described in any one of claims 1 to 6 when executing the computer program stored in the memory.

8. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which, when executed by a processor, implements the steps of data migration of the underwater temperature field satellite remote sensing prediction and reconstruction method based on the Transformer architecture as described in any one of claims 1 to 6.

9. A computer program product, characterized in that The invention comprises computer executable instructions, which, when executed, are used to implement the underwater temperature field satellite remote sensing prediction and reconstruction method based on the Transformer architecture as described in any one of claims 1 to 6.

Images