CN110516856B - A Convolutional Neural Network-Based Approach to Estimating Ocean Subsurface Temperature - Google Patents

Abstract

The invention relates to the technical field of marine science and technology, in particular to a method for estimating ocean subsurface temperature based on a convolutional neural network. In the method for estimating ocean subsurface temperature provided by the present invention, by obtaining observation data about ocean surface characteristics, and then extracting the data point with the highest surface correlation degree in the observation data, and performing convolution calculation on this basis, finally the ocean The estimated value of the subsurface temperature improves the efficiency of obtaining the estimated value of the ocean surface layer temperature, thereby improving the accuracy of the estimated value of the ocean subsurface layer temperature.

Description

Method for Estimating Ocean Subsurface Temperature Based on Convolutional Neural Networks

technical field

The invention relates to the technical field of marine science and technology, in particular to a method for estimating ocean subsurface temperature based on a convolutional neural network.

Background technique

The oceans play an unparalleled role in storing water and heat in the global climate system. Accurately detecting and describing the subsurface thermal structure of the global ocean is an important research direction of ocean dynamics.

In detecting and describing the subsurface thermal structure of the global ocean, studying and accurately estimating ocean subsurface temperature is crucial for understanding the mechanisms and processes of the entire ocean and the entire Earth's climate system. The construction of the set and the analysis of ocean warming provide data support.

Contents of the invention

The object of the present invention is to provide a method for estimating ocean subsurface temperature based on a convolutional neural network, said method comprising the steps of:

S1: Obtain observational data on surface characteristics of the ocean surface;

S2: Obtain the data point and the central data point with the highest surface correlation degree in the observation data;

S3: Carrying out a convolution budget on the data points to obtain the temperature representing the subsurface layer of the ocean.

Further, step S1 includes:

Obtain satellite observations of sea surface height, sea surface temperature, and sea surface salinity.

Further, the observation data is satellite observation data obtained by observing the ocean surface through remote sensing satellites.

In the method for estimating ocean subsurface temperature based on convolutional neural network provided by the present invention, by obtaining observation data about ocean surface characteristics, and then extracting the data point and central data point with the highest degree of surface correlation in the observation data, and using The convolution calculation is performed to finally obtain the estimated temperature of the ocean subsurface, which improves the acquisition efficiency of the estimated ocean surface temperature, and further improves the accuracy of the estimated ocean subsurface temperature.

Description of drawings

Fig. 1 is a flow chart of method steps provided by an embodiment of the present invention;

Fig. 2 is the CNN network structure that the embodiment of the present invention provides;

Fig. 3a and Fig. 3b are the comparison charts of measured values and predicted values at different depths of randomly selected near-coast points and far-coast points respectively provided by the embodiment of the present invention;

Fig. 4a and Fig. 4b are El Niño regional analysis diagrams provided by the embodiment of the present invention

Fig. 5a and Fig. 5b are El Niño regional analysis diagrams provided by the embodiment of the present invention

Fig. 6 is a trend chart of model prediction capability provided by an embodiment of the present invention.

Detailed ways

From the above description, it can be seen that in the global climate system, the ocean plays an incomparable role in storing water and heat. Accurately detecting and describing the subsurface thermal structure of the global ocean is an important research direction of ocean dynamics.

In detecting and describing the subsurface thermal structure of the global ocean, studying and accurately estimating ocean subsurface temperature is crucial for understanding the mechanisms and processes of the entire ocean and the entire Earth's climate system. The construction of the set and the analysis of ocean warming provide data support.

However, remote sensing observation data of the ocean surface are more frequently used to estimate the dynamic environment information of the ocean interior.

According to the research, the thermocline in the equatorial Pacific will remain abnormally deepened for a long time, which will cause the warming of the eastern equatorial Pacific to be greater than that of the region outside the equator and the weakening of the sea temperature gradient, which is exactly what leads to the frequent occurrence of extreme El Niño events One of the main mechanisms. The Intergovernmental Panel on Climate Change (IPCC) reported that the global average sea surface temperature (Sea Surface Temperature, SST) will increase by about 0.2°C per decade. Therefore, determining the SST and subsurface temperature (Subsurface Temperature, ST) is one of the most important issues in climate science for the study of the El Niño phenomenon and how it responds to greenhouse warming.

Satellite remote sensing provides many useful ocean surface observations on different spatial and temporal scales, but this is limited to the surface layer of the ocean. Since many important ocean processes and features are located below the sea surface and quite deep, and the existing data cannot completely and accurately describe the internal structure and changing laws of the ocean, it is very necessary to construct a complete three-dimensional temperature-salinity structure . With the continuous development of satellite remote sensing technology, especially the increasing data of sea surface temperature (SST) and salinity (SSS) from satellite remote sensing, a large number of satellites with wide coverage, high accuracy and spatial resolution are provided. High, time-continuous real-time sea surface information. How to use the data obtained by satellite remote sensing to predict ocean subsurface information and establish a complete set of subsurface three-dimensional analysis and prediction system is an urgent problem in the field of international ocean research.

In the prior art, Artificial Neural Network (ANN) is used to estimate sea surface temperature (SST), sea surface height (SSH), wind stress, net radiation flux and net heat flux data of Arabian Sea mooring system Ocean interior temperature structure. On average, 50% of the estimates were within ±0.5°C and 90% within ±1.0°C. For the temperature field at a depth of 200 meters in the North Atlantic Ocean, there is another existing technology to estimate the three-dimensional temperature field of the ocean through multiple linear regression combined with the ocean observation system data set Argo and remote sensing data, and use objective analysis methods. When combining the two data types, The mean square error of the mapping error for large-scale and low-frequency temperature fields at a depth of 200 m is reduced.

In addition, a new empirical method for estimating mesoscale three-dimensional ocean thermal structure from near-real-time satellite altimetry data is proposed in the prior art. The method uses a two-layer model with a stratified new set of empirical parameters. The researchers used sea surface temperature (SST), height (SSH) and salinity (SSS) data from the Argo grid monthly anomaly dataset to estimate the North Atlantic subsurface temperature anomaly using a self-organizing neural network. The depth of has good performance, and the correlation coefficient is greater than 0.8.

There are also researchers who predict ocean temperature by combining numerical values and artificial neural network technology. The error statistics of daily prediction are correlation coefficient r=0.37, root mean square error Root Mean Square Error RMSE=0.47℃ and mean absolute error Mean Absolute Error MAE=0.38°C, error statistics of weekly forecast are r=0.27, RMSE=0.78°C and MAE=0.64°C, error statistics of monthly forecast are r=0.11, RMSE=0.58°C and MAE=0.46°C. A new satellite-based geographically weighted regression (GWR) model developed by researchers in related fields is used to invert the subsurface temperature structure of the Indian Ocean. The final experimental results show that the RMSE range is about 0.10 to 0.18, and the R2 range is About 0.50 to 0.80.

Researchers estimate the subsurface temperature anomalies (Subsurface Temperature Anomalies, STA) of the Indian Ocean through a series of satellite remote sensing measurements, and propose a method of Support Vector Machines (SVM), and support vector regression (Support Vector Regression, SVR) The performance is unstable in the upper part of 400m, and with the decrease of ^R2 and the increase of Mean Squared Error (Mean Squared Error, MSE), the estimation accuracy gradually decreases with the depth of 500m.

We find that the large-scale SST annual cycle in the eastern equatorial Pacific is largely controlled by the annual variation of the mixed layer depth, which in turn is mainly determined by the competing effects of solar radiation and wind forcing. The SST in the tropical and temperate Pacific Ocean was abnormally unstable in spring. This instability is crucial for the presence of predictability barriers to El Niño events.

Under the influence of global warming, the average climate in the Pacific region may change significantly. In the past, the data prediction models established by researchers on an annual basis were affected by climate and seasonal changes. These data prediction models usually need to subtract the climate mean state to eliminate the influence on the training model, but the results show that the STA prediction ability is insufficient. This study proposes to establish a monthly data model to eliminate the impact of climate on model training. The results show a significant improvement in STA prediction accuracy.

In the past, researchers only used single-point features to establish STA prediction models, which did not consider the influence of seawater flow or ocean currents on seawater heat transfer in the ocean. Its results showed that the prediction accuracy of STA was poor.

Based on the above analysis, it can be seen that existing methods for retrieving subsurface thermal structure from sea surface parameters are usually based on dynamical or statistical models. Existing dynamical methods pay little attention to global-scale applications and advanced machine learning models, and use only a small number of surface parameters to derive subsurface dynamical fields. Therefore, the estimation method itself and its global scale accuracy still show a lot of room for improvement. Existing statistical methods are not adequately applied to ocean surface data parameters, lack of application of deep learning models, single input characteristic parameters of the model, poor learning ability, and the establishment of models based on annual data is greatly affected by climate. Therefore, the prediction accuracy of the model has not been significantly improved.

The method for estimating ocean subsurface temperature based on the convolutional neural network proposed by the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer from the following description. It should be noted that all the drawings are in a very simplified form and use imprecise scales, and are only used to facilitate and clearly assist the purpose of illustrating the embodiments of the present invention.

In the detailed description of the present embodiment, spatial terms such as "below", "beneath", "below", "above" and the like are used for the purpose of easily describing one part and the other shown in the drawings. positional relationship of one component, but these are only examples and are not intended to limit the present invention. The spatially relative terms are intended to encompass various orientations of the device in use or operation in addition to the orientation depicted in the figures. A device may be otherwise oriented, eg, rotated 90 degrees or at other orientations, and interpreted accordingly by the spatially relative descriptors used herein.

This example proposes a method for estimating ocean subsurface temperature based on a convolutional neural network. The method comprises the steps of:

S1: Obtain observational data on surface characteristics of the ocean surface;

S2: Obtain the data point and the central data point with the highest surface correlation degree in the observation data;

S3: Perform a convolution operation on the data points to obtain an estimated temperature of the ocean subsurface.

Further, step S1 includes:

Obtain satellite observations of sea surface height, sea surface temperature, and sea surface salinity.

Further, the observation data is satellite observation data obtained by observing the ocean surface through remote sensing satellites.

Specifically, this embodiment uses data sets from 2004 to 2015 for model training and evaluation, in which data sets from 2005 to 2014 are used for model building, and data sets from 2004 and 2015 are used for Model evaluation. The data sets from 2005 to 2014 were divided into 12 different data sets by month, and 12 groups of monthly CNN models were established. Each set of data samples is 140,000, of which 78,400 (56%) data are used to train the data model, 19,600 (14%) are used to verify the model, and 42,000 (30%) are used to test the model.

The estimation method in this embodiment is first to standardize the ocean surface features, which ensures that the dimensions and magnitudes of different features have the same value range. When processing, the data feature value of a single sample is subtracted from the mean value of the same feature of all data in the training sample, and then divided by its variance. In this way, for each feature, all data are clustered around 0 with a variance of 1. The specific publicity calculation is as follows:

x is the eigenvalue of a single sample in the training set or test set, μ is the average value of the training sample data, σ is the standard deviation of the training sample data, and X is the eigenvalue after normalization.

The scheme disclosed in this embodiment associates the temperature data of the subsurface layer of the central point of Argo with the corresponding position of the CMEMS data according to the coordinate values of longitude and latitude, so as to eliminate the impact of the mismatch between CMEMS and Argo resolution (the resolution of CMEMS is 0.25° ×0.25°, Argo resolution is 1°×1°, uniformly 0.25°×0.25°).

The training eigenvalues selected in this study are the center point and surrounding data points (a total of 625 data points, and the number of eigenvalues is 1875).

In addition, convolutional neural networks (CNN) have a more complex network structure and have stronger feature learning and feature expression capabilities than traditional machine learning methods. Each neuron is regarded as a filter (filter), its window (receptive field) slides, and the filter (filter) calculates the local data. The Relu excitation layer performs nonlinear mapping on the output of the convolutional layer. The pooling layer is sandwiched between consecutive convolutional layers, which is used to compress the amount of data and parameters and reduce overfitting. Usually the fully connected layer is at the end of the convolutional neural network, which is the same as the connection method of traditional neural network neurons.

The advantage of CNN lies in the shared convolution kernel, which is highly efficient for processing high-dimensional data, and can automatically extract some advanced features, reducing the time of feature engineering, which can improve the accuracy of predicting subsurface temperature. Therefore, it is practical and feasible to use the convolutional neural network (CNN) to build a model using the data features with the highest degree of surface correlation.

Figure 2 shows the set CNN network structure. The structure consists of 5 layers of convolution operations, followed by 4 layers of fully connected layers.

The data point with the highest correlation degree on the surface layer and the central data point are regarded as a two-dimensional image. The image is subjected to the convolution operation of the local field of view, and the three features (SSTA, SSHA, SSSA) of each data point are operated as a convolution unit. Use RELU activation function and Adam optimization function.

The stride of the convolutional layer of the first layer is (3,1), that is, it moves 3 steps in the horizontal direction and 1 step in the vertical direction. The size of the filter filter is 3×2, that is, the corresponding elements of the features of the two data points are multiplied by the corresponding elements of the filter and then summed. After calculating one area, move the specified stride (3,1) to other areas until the 57× The two-dimensional matrix of 33 is completely covered. After 5 layers of convolution operations, the data dimension is flattened into one-dimensional data, input to the fully connected layer, and then 4 layers of neural network operations are performed, and finally 57 layers of predicted values are output.

The process of estimating the Pacific Subsurface Temperature Anomaly (STA) by establishing a monthly CNN model based on multi-source remote sensing observation data (SSTA, SSHA, SSSA) of the sea surface (taking 100m depth as an example).

First, build the training dataset. The selected training eigenvalues are the data points with the highest correlation between the center point and its surface (a total of 625 data points, and the number of eigenvalues is 1875). All datasets are standardized using Argo's measured STA as training marks and test marks.

Second, train the CNN model and build the optimal CNN model. The model chooses Relu as the activation function and Adam as the optimization function. We determine the optimal combination of convolutional layers, pooling layers, and fully connected layers by analyzing each MSE, ^R2 , and convergence speed. We use the training datasets (SSTAx, SSHAx, SSSAx) as input data for CNN training and Argo's STA as training markers.

Finally, we use the CMEMS datasets (SSTA, SSHA, SSSA) as the input parameters of the CNN model to predict the subsurface STA. We use the STA measured by Argo to evaluate the prediction accuracy of the CNN model at each layer (57 layers) of the subsurface.

The present embodiment also provides the analysis to above-mentioned estimation temperature, and Fig. 3 a and Fig. 3 b show this experiment randomly selected point a (21.50 ° N, 122.50 ° E) and point b (14.50 ° N, 160.50 ° E) ( Comparison chart of measured and predicted values at different depths (October 2015) (the MSEs of points a and b are 0.1980/0.0980, respectively). The results show that the method we propose to use the characteristics of the data points with the highest surface correlation to predict the center point not only has a good prediction effect on the regular data points at the far coast, but also has a good prediction effect on the irregular data points near the coast. . This shows that the method also generalizes to irregular data points near the coast.

Figure 4a, Figure 4b, Figure 5a, and Figure 5b show the content for El Niño regional analysis. In the future, we hope to further study the abnormal year information by collecting more data sets of abnormal years.

The low-temperature water in the tropical Pacific Ocean is obvious at different depths, especially at the depth of 100m to 300m, which is related to the La Niña phenomenon in the upper layer of the tropical Pacific Ocean in March and April. The STA at different depths in the tropical eastern Pacific Ocean was significantly lower than in other sea areas. As the depth increases, the seawater temperature generally tends to be stable, the intensity of STA changes becomes smaller and smaller, and the spatial heterogeneity becomes less obvious, which is related to the difference in dynamics between the interior and surface of the ocean.

This embodiment uses MSE and ^R2 as model evaluation indicators. We only select the months (1, 4, 7, 10) with obvious ocean climate change as the analysis data. From Table 1 and Table 2 and Figure 4a, Figure 4b, Figure 5a, and Figure 5b, it can be obtained that in January, April, July, and October of 2004 and 2015, the MSE was the highest when the depth horizon was between 30m and 200m (the highest MSE value 0.8408/1.1034/1.8640/2.8231/0.5961/0.6750/1.1856/0.4788), which may be related to the complex dynamic process in the upper Pacific Ocean and the disturbance of the mixed layer and thermocline.

Table 1

Table 1 Comparison of mean square error (MSE) and coefficient of determination (R ² ) corresponding to CNN models at different depths and months in 2015.

Table 2

Table 2 shows the mean square error (MSE) and coefficient of determination (R ² ) corresponding to the CNN model at different depths and in different months in 2004.

As shown in Figure 6, it shows that after the depth of 300m, the MSE tends to be stable, ^R2 gradually decreases, and the prediction ability of the model decreases. This may be due to the relatively stable state of stratification in the middle and deep seawater, and the deeper the physical phenomenon in the ocean, the more difficult it is to use. Surface features are predicted and harder to detect by satellite remote sensing. From the table, it can be seen that the prediction accuracy of the model varies in different years and months:

(1) In the same year, taking the depth of 100m in 2015 as an example, the average MSE in January was 0.2821 (the maximum/minimum value was 0.8408/0.0062), and the average ^R2 was 0.966 (the maximum/minimum value was 0.993/0.891). The MSE (R ² ) in January was lower (higher) than that in April, July, and October, and the prediction accuracy of the model has improved, which is related to the fact that the seawater temperature tends to be stable in winter and the intensity of STA changes is small. In July and October, ^due to the obvious changes in the internal temperature of seawater in summer and autumn, the Pacific Ocean is strongly affected by ocean currents, MSE increases and R ² decreases. The maximum/minimum value is 0.993/0.842), the average MSE in October is 0.8027 (the maximum/minimum value is 2.8232/0.0034), and the average ^R2 is 0.956 (the maximum/minimum value is 0.983/0.939). The above shows that the predictive ability of the CNN model decreased in July and October, mainly due to the drastic changes in the internal temperature of seawater in summer and autumn, and the influence of ocean currents.

(2) In different years, the MSE at different depths in different months in 2004 was generally low (the average MSE in 2004 was 0.2859, and the average MSE in 2015 was 0.5117). In 2015, R ² was generally higher than (the average R ² in 2004 was 0.969 in 2015 and the average ^R2 was 0.958) in 2015. The prediction results of the model in January, April, July, and October are more reliable for 2004, but decreased in 2015. This is mainly due to the fluctuation of climate change in recent years, the temperature anomaly in the Pacific Ocean is obvious, and the intensification of uncertain factors leads to a significant decline in the prediction accuracy of the model.

In summary, in the method for estimating ocean subsurface temperature provided by the present invention, by obtaining observation data about ocean surface characteristics, and then extracting the data point and central data point with the highest degree of surface correlation in the observation data, and using this The convolution calculation is performed to finally obtain the estimated value of the ocean subsurface temperature, which improves the acquisition efficiency of the estimated ocean surface temperature, and further improves the accuracy of the estimated ocean subsurface temperature.

The above description is only a description of the preferred embodiments of the present invention, and does not limit the scope of the present invention. Any changes and modifications made by those of ordinary skill in the field of the present invention based on the above disclosures shall fall within the protection scope of the claims.

Claims (2)

1. a method for estimating ocean subsurface temperature based on convolutional neural network, is characterized in that, described method comprises the steps: S1: Obtain observational data about the surface layer characteristics of the ocean surface, the observational data refers to satellite observational data including sea surface height, sea surface temperature and sea surface salinity; S2: Obtain the data point with the highest surface correlation degree and the central data point in the observation data, and regard the data point with the highest surface correlation degree and the central data point as a two-dimensional image; S3: Perform a convolution operation on the data points; wherein the convolution operation includes: First, construct the training data set, select the training feature value as the center point and the data point with the highest surface correlation degree, use Argo to measure the subsurface temperature anomaly as the training mark and test mark, and all data sets are standardized; Second, train the CNN model and build the optimal CNN model. The model selects Relu as the activation function and Adam as the optimization function. By analyzing each MSE, R ² and convergence speed, the parameters of the convolutional layer, pooling layer, and fully connected layer are determined. optimal combination; Finally, using the CMEMS data set as the input parameter of the CNN model, the subsurface temperature anomaly is predicted; The normalization process includes associating the subsurface temperature data of the central point of Argo with the corresponding position of the CMEMS data according to the coordinate values of longitude and latitude. 2. The method for estimating ocean subsurface temperature based on convolutional neural network as claimed in claim 1, wherein the observation data is satellite observation data obtained by observing the ocean surface through remote sensing satellites.

Images