CN114139590A

CN114139590A - Method for estimating ocean temperature

Info

Publication number: CN114139590A
Application number: CN202110587615.9A
Authority: CN
Inventors: 冯源; 孙天颖; 李晨
Original assignee: Ocean University of China
Current assignee: Ocean University of China
Priority date: 2021-05-27
Filing date: 2021-05-27
Publication date: 2022-03-04

Abstract

The disclosure relates to the technical field of marine science, in particular to a method for predicting marine temperature change by using a deep learning network model. According to the method for estimating the ocean temperature, the data acquired in different ways are assimilated, so that more ocean meteorological data with larger quantity and larger scale can be acquired, a data set with enough quantity and range for training the model can be constructed, and a more accurate estimation model can be acquired.

Description

Method for estimating ocean temperature

Technical Field

The disclosure relates to the technical field of marine science, in particular to a method for predicting marine temperature change by using a deep learning network model.

Background

Sea surface temperature has a major impact on the health of regional marine ecosystems, and its trend may lead to the growth, reproduction and distribution range of marine species. The ultra-long-term sea surface temperature prediction in large-scale sea areas has important guiding significance on large-scale marine physical phenomena, and is beneficial to the work in the aspects of climate monitoring, flood and drought risk early warning and the like.

Under the global warming environment, the global sea surface temperature is increased to a certain extent, and the local climate and the marine ecosystem are influenced significantly. The existing sea surface temperature forecast provides effective data reference for the research on sea surface temperature evolution characteristics, climatic phenomena, weather changes and the like. The current research adopts models such as Artificial Neural Network (ANN) and Support Vector Machine (SVM), and mainly focuses on short-period prediction of medium and small-scale sea areas.

Obtaining accurate large-scale, long-term surface temperature predictions remains a significant challenge.

Disclosure of Invention

In order to solve the above technical problem, the present disclosure provides a method capable of estimating ocean temperature, comprising the steps of:

assimilating the marine meteorological data with different acquisition ways and preset resolution to obtain an assimilation data set

And inputting the assimilation data set into a preset network model for training to obtain an estimation model.

Further, the estimation model is a convolution network model.

Further, after the convolutional network model is subjected to dilation convolution, the size formula of a convolution kernel is as follows:

K＝(k-1)*d+1

where K is the size of the expanded convolution kernel;

k is the size of the original convolution kernel;

d is the expansion rate of the neural network layer in the convolutional network model.

Further, the oceanographic data at least comprises data information of 8 months.

According to the method for estimating the ocean temperature, the data acquired in different ways are assimilated, so that more ocean meteorological data with larger quantity and larger scale can be acquired, a data set with enough quantity and range for training the model can be constructed, and a more accurate estimation model can be acquired.

Drawings

FIG. 1 is a schematic diagram of the TCN structure provided by the present invention;

FIG. 2 is a schematic diagram of the density of data of different correlation degrees according to the present invention;

FIG. 3(a), FIG. 3(b), FIG. 3(c), FIG. 3(d), FIG. 3(e) show observed and predicted SSTS profiles;

FIGS. 4(a) and 4(b) are schematic diagrams showing the spatial distribution of the correlation between the model prediction and the true value;

FIG. 5(a) shows an ACC line graph for each month in 2014-2018;

FIG. 5(b) is a RMSE line graph of months 2014 to 2018.

Detailed Description

At present, common ocean temperature estimation usually adopts models such as Artificial Neural Networks (ANNs) and Support Vector Machines (SVM), and mainly focuses on short-period prediction of medium and small-scale sea areas. It is difficult to make long-term temperature estimation predictions. Time-spanning ocean temperature estimation is also considered to be an industry challenge for workers.

In order to solve the difficulties faced by the prior art, the method utilizes observation data with a span of 10 years to predict the monthly average ocean temperature of the next 5 years, wherein the forecasting work with the span of 5 years can be realized only by using one model, the monthly average value of Root Mean Square Error (RMSE) is 0.506 ℃, and the sea surface temperatures are arranged into SSTs time sequence according to the chronological order. The average correlation between the observed and predicted time series of SSTs was 88.23%. Compared with a climate forecast system V2(CFSv2) published by the National Oceanic and Atmospheric Administration (NOAA) in 2010, which needs to continuously collect observation data and uses 7 models to complete 5-year prediction work, the model training difficulty and the number of models are greatly reduced by the method.

The method takes the space-time autocorrelation of the sea surface temperature into consideration, selects the space-time data with high correlation degree as the characteristics, inputs the characteristics into a model in a numerical form, and extracts the space-time information. In the past, the spatial dependency relationship is learned by inputting a two-dimensional matrix image into a model, and the image processing is complex and the calculation time is long. The TCN method is adopted to directly process the numerical value, so that the time for generating the model is greatly shortened, and the performance requirement on hardware equipment is reduced.

The method provided by the disclosure can effectively improve the stability of the model and has profound influence on the research of the medium-term and long-term ocean phenomenon. The change under the sea is likely to leave traces on the sea surface through the change of the sea level height, and the research on the sea surface temperature has important significance on subsurface parameters.

The method adopts a time domain convolution network (TCN) to utilize multi-source multi-modal sea air data for modeling, sea surface temperature forecast of 5-year period is carried out on sea areas (the spatial resolution is 1 degree multiplied by 1 degree) of 40 degrees E-110 degrees E, 25 degrees N-25 degrees S, and ultra-long period sea surface temperature forecast of large-scale sea areas is completed. At the same time, the present disclosure combines some theories of marine physics with deep learning. The ocean surface is subject to ocean currents and turbulence. The annual period of large SSTs in the east of the equatorial pacific is to a large extent governed by the annually varying depth of the hybrid layer, which in turn depends mainly on competing effects of solar radiation and wind forces. Therefore, the method is characterized by selecting multi-factor and multi-level data such as ocean surface, sea, atmosphere and the like. The TCN model contains prior knowledge of physical oceanography in the training process, and the aim of improving the prediction precision of the Sea Surface Temperature (SST) is fulfilled by the strong mining capability of the deep learning model on data. (Red-New)

The research of the physical ocean phenomenon in the sea area needs to process huge and fine data, the method provided by the disclosure uses a data set with low spatial resolution and small time granularity to predict the ultra-long time period, and the prediction precision is ensured while the area covering the sea area is large. The method can be applied to the actual prediction of large marine physical phenomena such as sea surface detection, vortex recognition and the like.

2.1Data

Specifically, the present disclosure is directed to the interior of the indian ocean, which is the third major ocean of the world (30 ° E-135 ° E, 30 ° N-66.5 ° S). The Indian ocean affects the climate abnormalities in the peripheral regions, the central region of south America, the southern end of Africa, the southeast Australia, and northeast Asia. The present disclosure uses a re-analysis data set with a spatial resolution of 1 ° × 1 ° provided by the american meteorological environment prediction center (NCEP) to perform quality control and assimilation processing on observations from various sources (ground, ships, radiosondes, windfinding balloons, airplanes, satellites, etc.), where the assimilation processing represents the continuous fusion of direct or indirect observation information from different sources and different resolutions, discretely distributed in space and time, by a data assimilation algorithm. The data set has the characteristics of multiple contained elements, wide range and long extension time. The present disclosure selects oceanic-atmospheric monthly data for the longitude (40 ° E-110 ° E) latitude (-25 ° S-25 ° N) sea area in the indian ocean. The sea temperature prediction problem can be solved as a time series regression problem, and the method aims at solving the problems that in the prior art, researchers only use sea surface temperature to construct a time series, train a surface temperature prediction model and do not consider sea water in the sea to construct a time series, and do not consider the sea surface temperature prediction model

The influence of flow or ocean current on the heat transfer of the seawater leads to the problem of insufficient prediction effect of the surface sea surface temperature, and the influence of various factors on the surface sea temperature is fully considered when the characteristics are selected.

TABLE 1 81 characteristics affecting the temperature of a sea surface

The present disclosure selects monthly data for successive decades, predicting the sea surface temperature for the next 5 years. Table 1 shows that the possible factors (81 in total) affecting the data point and the sea surface temperature are also data of each month, and are divided into three parts, namely atmospheric, sea surface and undersea parameters. The present disclosure selects atmospheric temperature (T), potential altitude (GPH), vertical velocity (W), Relative Humidity (RH), east-west wind velocity (U), north-south wind velocity (V), data at different altitudes (1000850500300hPa), for a total of 24 atmospheric factors. The sea surface parameters include Sea Surface Height (SSH) at the center point, Sea Surface Temperature (SST) at the center point, and Sea Surface Temperature (SST) at 15 points around (indicated by sub1_ SST, sub2_ SST.. sub15_ SST in the table), which are 17 in total. The sub-sea parameters comprise the temperature at different depths of the sea (5, 15, 25, 35, 45, 55, 65, 75, 85, 95m) of the centre point and 40 for east-west u and north-south v, salinity (SSS). The training data set comprises 20 × 2533 pieces of data, namely the number m of samples, and the testing data set comprises 2533 pieces of data to form oceanographic data with different acquisition ways and preset resolution. The present disclosure sorts the training set in a sliding window fashion. The training set features comprise data from 1980 to 2008, the feature time span is ten years, and the sliding window slides backwards one year at a time to form a training set, namely a marine meteorological data set for training a model. The test set (i.e. the oceanographic data set for testing the model accuracy) comprises 2533 pieces of data of 2004-2013 ocean atmospheric month data and corresponding month sea surface temperatures of 2014-2018.

Then, deep learning is carried out on the ocean big data by adopting a time domain convolution network (TCN) to obtain an estimation model.

The present disclosure uses the TCN architecture. Because the TCN applies residual connection, expansion convolution, causal convolution and other ideas, the TCN structure has more advantages when the problem of long time sequence space is processed. Residual concatenation (i.e., a residual block) can eliminate to some extent the effects of gradient vanishing and explosion of the deep network portion. Causal convolution enables prediction y at time t by limiting the sliding direction of the convolution window_tOnly by input x before time t₁To x_t-1And then the judgment is carried out. The expansion is originally applied to the field of image segmentation, and the lost information is reduced and the receptive field is increased while the input dimension and the output dimension are kept the same. The following is a formula of the size of the convolution kernel after the expansion convolution, K is the size of the convolution kernel after the expansion convolution, K is the size of the original convolution kernel, and d is the expansion rate of the neural network layer:

k ═ K-1 × d +1 formula 1

The present disclosure processes a two-dimensional training set into an (m, 120, 81) three-dimensional matrix input model, where each sample is a matrix with a size of (120, 81), 120 represents a time step, 81 represents a feature at one time step, the number of layers of the three-dimensional matrix is m, and m is the number of samples. The number of convolution kernels is set to be 8 and 24 respectively, relationships are [ 1, 2, 4, 8, 16, 32, 64, 128, 256 ], and Stack is 1. Experiments have shown that the model works best when the size of the convolution kernel is set to 8, which means that the model is trained with information of the last eight months being considered at the shortest, and the history time is considered longer and longer as the expansion factor increases. The number of convolution kernels determines the number of feature maps generated in the convolution, the feature maps contain information extracted from the previous layer of output, and the information extracted by different convolution kernels is different. When the number of convolution kernels is too large, some tiny and accidental data disturbance in a training set can be learned by the model, and therefore the accuracy of the model is affected. When the number of convolution kernels is too small, the model learning characteristic capability is weak. The present disclosure makes the output information contain more history data by setting relationships lists. When relationships are set to 1, the model is trained in the manner of normal convolution.

Fig. 1 shows the operation of a sample in a TCN structure. The network architecture used by the present disclosure includes 28 layers of convolution operations, a 1 layer of scatter layers, and a 1 layer of fully-connected layers. After a one-dimensional convolution (Conv1d) is performed on the samples of the input layer, a matrix with a size of (120, 19) is output and enters the first residual block for operation (red box). The TCN model of the present disclosure sets nine residual blocks. The non-linear mapping result of each residual block is retained and added to the convolution result of the last residual block. And finally, mapping a prediction result (60, 1) by a Flatten layer and a full-link layer to form a diagram, wherein two area frames are used for distinguishing convolution operation and full-link operation, the area frame on the left is used for 28-layer convolution operation, and the area on the right is used for crazing the Flatten layer and one layer of full link.

Taking 2014-2018 prediction models as examples, the training set of the input models is a matrix (120, 81, 50660), 20% of the training set is used as a verification set, and model parameters are adjusted by the verification set to obtain an optimal model, namely an estimation model which is finally required to be obtained by the present disclosure. The model predicts a lunar average sea surface temperature (60) for valid data points (2533) on the indian ocean basin, with an output matrix size of (60, 2533). The method adopts the correlation, the RMSE, the ACC and the performance of a single-month prediction value and an observation value of the model on an integral data set to evaluate the accuracy of the TCN model in predicting the average sea table temperature in 2014-2018 months. The correlation profile was drawn by calculating the correlation of the 5-year true ssts time series for each data point and the model estimated ssts time series. The higher the correlation, the more similar the sea surface temperature sequence trends, and the higher the model fitness. Where ACC represents the average prediction accuracy of the model over the data points.

From fig. 2, it can be seen that the data points with the correlation degree between 0.90 and 0.97 are dense, the number of data points in the section with the correlation degree below 0.75 is less than 50, and it can be seen that only a few data points have the correlation degree less than 0.75. The correlation degrees of the SSTs sequence observed by 2533 effective data points in the Indian ocean and the SSTs sequence estimated by the model show significant positive correlation, wherein the correlation degrees are both more than 55 percent, the data points with the correlation degree within the range of 55 to 65 percent only account for 1.54 percent of the total number of the data points (2533), and the correlation degree of nearly half of the total number of the data points is 85 to 95 percent. The correlation degree of the data points is basically over 75 percent. The average correlation of the indian ocean basin data points was 88.23%. The degree of fit of the model to the data points is overall higher.

And randomly selecting five data points from the five correlation degree intervals to show the true sea temperature and the predicted sea temperature change in 2014-2018. The model can approximately fit the sea temperature change trends of the data points (1) and (2) and learn the average trend of the sea temperature change, and the data points (3) to (5) observe the abnormal rise of the true sea temperature in the first half of 2016, so that the prediction error of the sea surface temperature in the first half of 2016 is larger at the data points (3) to (5) because the model tends to learn an average general change trend, and the relevance of the data points is low.

Selecting 1 data point from the five relevance interval ranges, and showing the observed ssts sequence curve and the predicted ssts curve of 5 data points in FIG. 3, the closer the tendency of the ssts sequence curve is, the higher the relevance is. FIGS. 3(a), (b) estimate that the tendency of changes in the ssts time series versus the actual ssts series is highly fitting and highly correlated. The data points in fig. 3(c), (d), and (e) show abnormal changes in sea surface temperature in months 6-2016, and the error between the prediction and the true sea temperature is large, resulting in a decrease in the correlation between the data points.

Fig. 5(a) TCN model predicts SSTs time series and true SSTs sequence correlation, and fig. 5(b) CFSv2 model predicts SSTs time series and true SSTs sequence correlation as used in this disclosure. Overall, the TCN model of the present disclosure is more highly correlated, with lower correlation data points concentrated in the sea area from south to-10 ° S, 75 ° E to 90 ° E, above the equator, and the CSFv2 model shows lower correlation data points in many regions.

FIG. 4(a) shows the spatial distribution of TCN model predicted correlation, with 55% -65% of the data points centered between-7 ℃ S and-2 ℃ S, and 75 ℃ E and 100 ℃ E. The correlation of the data points located in the middle of the sea area under study ranged from 66% to 75. In the low-latitude areas of the east coast of the Indian ocean, the degree of correlation is also reduced to 75-85%. The correlation degree of coastal areas between the Somari peninsula and Arabic and Indian peninsula and western areas, particularly southwest sea areas, is higher and is more than 85 percent. Overall, the fitting degree of the Indian ocean basin ssts time sequence is more than 75%, and the average correlation degree is 88.23%, which proves that the model has good robustness.

Fig. 4(b) shows correlation distribution between real values and model predicted values in 2014-2018 of a climate forecasting system V2(CFSv2) released by the National Oceanic and Atmospheric Administration (NOAA) in 2010. The CFSv2 model only makes periodic 9-month predictions of the global ocean surface temperature. So the present disclosure selects one model every 9 months for a total of 7 predictive models of NOAA. The predicted values (the resolution is 0.9375 degrees x 0.9375 degrees) of the CSFv2 model at-25.039 degrees S-25.039 degrees N and 40.312 degrees E-109.687 degrees E are selected and compared with the true values (1 degree x 1 degrees). The data points with the longitudes of 54.375, 71.250, 88.125 and 105 and the latitudes of 16.535, -0.472 and-16.535 are selected by the method and deleted so that the majority of data points in the NOAA are close to the longitude and latitude of the real data point.

The average correlation of the CFSv2 prediction model is 87.27%, and the average correlation of the TCN model of the disclosure is 88.23%. The correlation degree of individual points of the CFSv2 model is lower than 0.55, and the correlation degree of the model of the present disclosure is higher than 0.55. As can be seen from the figure, the correlation degree of the equator north is higher than that of the CFSv2 model in the whole, and the prediction correlation degree of the model is high, so that the prediction capability of the model on the change trend of the sea surface temperature time series is stronger. Multiple CFSv2 models are needed to complete the long-period forecast, and the period forecast of 5 years is completed by only one model, so that the method is more functional than the CFSv2 model.

FIG. 5(a) is an ACC line graph for each month in 2014-2018, and (b) is a RMSE line graph for each month in 2014-2018. ACC from 8 to 2016 was less than the average ACC value and a significant increase in RMSE error occurred in agreement with the RMSE error plot.

FIG. 5 is a comparison between the RMSE and the ACC of the 2014-2018 prediction model, wherein the average value of the ACC is 0.985, and the average value of the RMSE error is 0.506 ℃. RMSE fluctuation is obvious from 8 months to 2016 months, with large error and low ACC. This is consistent with the periods of greater deviation of the sea surface temperature predictions shown in fig. 3(c) - (e). In other years, the error is stabilized between 0.3 ℃ and 0.5 ℃.

The method carries out sea surface temperature prediction of a large-scale time period in a large-scale sea area, estimates the sea surface temperature of five years in the future by utilizing sea and atmosphere observation data of 10 years, and estimates the model by predicting SSTs time sequence correlation degree, RMSE and ACC. The average monthly error of the sea surface temperature of 2014-plus 2018 is predicted by the method is 0.506 ℃.

The current research focuses on the small-scale and medium-scale sea surface temperature forecasting in a small-scale sea area, and the data set resolution is high, while the data set resolution is lower but the accuracy is higher for the forecasting of a large-scale time period in a large-scale sea area.

The model of the present disclosure has important significance for the forecast of marine physical phenomena and is also reflected in the aspect of feature selection. Large-scale oceanographic phenomena are influenced by multiple factors such as ocean currents, sea surface wind, and light. The model disclosed by the invention combines some theories of marine physics with characteristic selection engineering, models multi-source multi-modal data, and learns the relationship between a plurality of characteristics and the sea surface temperature by learning factors such as the sea surface temperature, the flow velocity and the atmosphere, so that the model is more in line with the research requirements of physical phenomena. In addition, a convolutional network such as CNN is generally adopted when spatial position information is learned in the past, the model is input in a picture mode, the calculation amount is large, the TCN model disclosed by the invention is used for inputting numerical values, modeling is carried out, and the calculation is simpler and quicker. The ocean data system is numerous and complicated, and obviously, the model disclosed by the invention is more suitable for processing the ocean data. The model of the present disclosure outputs a five year sea surface temperature time sequence at a time, which is more functional than the CFSv2 model predicted nine months in advance by NOAA.

Claims

1. A method for estimating ocean temperature, comprising the steps of:

assimilating the oceanographic data with different acquisition ways and preset resolution to obtain an assimilation data set;

2. The method for estimating ocean temperature according to claim 1, wherein the estimation model is a convolutional network model.

3. The method for estimating ocean temperature according to claim 2, wherein the convolution network model, after performing the dilation convolution, has a convolution kernel size formula as follows:

K＝(k-1)*d+1

where K is the size of the expanded convolution kernel;

k is the size of the original convolution kernel;

4. The method for estimating ocean temperature of claim 1, wherein the metocean data comprises at least 8 months of data information.