A method for predicting high-resolution three-dimensional temperature and salinity field changes using multi-source ocean data

Abstract

The application discloses a method for predicting high-resolution three-dimensional temperature and salinity field changes by using multi-source marine data. The method converts three-dimensional temperature and salinity prediction into horizontal marine surface element prediction and joint inversion of the next marine subsurface temperature and salinity. Firstly, a sea surface temperature prediction model and a sea level anomaly prediction model are constructed by using a M-ViT model with low-resolution satellite remote sensing data. The predicted sea surface temperature and sea level anomaly prediction data are differentiated and matched with high-resolution subsurface three-dimensional temperature and salinity data. Then, a SimVP-gsta model is used to construct a subsurface three-dimensional temperature and salinity prediction model, so as to realize the prediction of high-resolution subsurface three-dimensional temperature and salinity by using low-resolution data of sea surface temperature and sea level anomaly. The application avoids the problem of large amount of data caused by stacking of high-resolution historical data, effectively increases the prediction time under the premise of not reducing the time and space resolution of the prediction result, and accurately predicts the marine temperature and salinity in most layers.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of ocean temperature and salinity field prediction, and specifically relates to a method for predicting high-resolution three-dimensional temperature and salinity field changes using multi-source ocean data. BACKGROUND

[0002] High-resolution three-dimensional temperature and salinity data can provide fine and comprehensive data support and analysis for ocean environment monitoring. These data not only help to better understand the physical and chemical processes within the ocean, but also provide important scientific basis for climate change research, marine ecosystem protection, and marine resource development. Through monitoring and analysis of high-resolution three-dimensional temperature and salinity fields, scientists can more accurately predict changes in ocean temperature and salinity fields, thereby improving the response capability to extreme weather events, marine disaster warning, and marine environmental protection. Therefore, improving the accuracy of ocean three-dimensional temperature and salinity field prediction has far-reaching significance for promoting marine scientific research and promoting marine environmental protection.

[0003] The current mainstream three-dimensional prediction method is based on traditional numerical model prediction. This method uses mathematical equations to describe ocean movement and physical processes, and solves through numerical simulation. However, the traditional numerical prediction method requires high computing resources and mainly relies on parameterized numerical models, which are affected by initial conditions and parameter selection and have uncertainties, resulting in larger errors in long-term prediction.

[0004] With the rapid development of data-driven deep learning methods, ocean prediction methods based on deep learning are considered as a powerful supplement to traditional methods. Related models of spatiotemporal sequence prediction based on deep learning are applied to the spatiotemporal prediction of sea surface temperature. These models can capture the spatiotemporal correlation and nonlinear characteristics of sea surface temperature data, learn from historical meteorological data, discover hidden patterns and rules in the climate system, and then predict the future trend of sea surface temperature changes. However, past studies on ocean temperature and salinity prediction have mostly focused on two-dimensional planes, but for ocean environmental monitoring, more attention should be paid to effective prediction in three-dimensional space. In recent years, Zuo et al. used a deep learning model to predict ocean temperature, but the prediction time was only one day, and the prediction area was small, which made it difficult to meet the needs of ocean environmental monitoring (Reference 1: Zuo X, Zhou X, Guo D, et al. Ocean temperature prediction based on stereo spatial and temporal 4-D convolution model [J]. IEEE Geoscience and Remote Sensing Letters, 2021, 19: 1-5.). Sun et al. used a 3D-UNet model to predict the sea surface temperature (SSbT) above 400 meters in the Pacific Ocean and its adjacent oceans, but the data was monthly average data in terms of time resolution, and the spatial resolution was only 1°x1° (Reference 2: Sun N, Zhou Z, Li Q, et al. Spatiotemporal prediction of monthly sea subsurface temperature fields using a 3D U-Net-Based model [J]. Remote Sensing, 2022, 14(19): 4890.). Wang et al. proposed the first global ocean high-resolution prediction model XiHe. Due to the stacking of three-dimensional high-resolution data in prediction time, it took three months to train on eight NVIDIA DAH100 (Reference 3: Wang X, Wang R, Hu N, et al. Xihe: A data-driven model for global ocean eddy-resolving forecasting [J]. arXiv preprint arXiv: 2402.02995, 2024.). SUMMARY

[0005] In view of the problems of short prediction time, low spatio-temporal resolution and difficult model training in the current deep learning-based marine prediction technology, the application provides a method for predicting high-resolution three-dimensional temperature and salinity field changes by using multi-source marine data, which uses low-resolution data on the sea surface for two-dimensional prediction, avoids the problem of stacking three-dimensional high-resolution data in the time dimension, reduces the difficulty of model training, saves a large amount of computing resources, and uses satellite remote sensing data for input to obtain high-resolution three-dimensional temperature and salinity prediction results, which is more practical in application.

[0006] The application provides a method for predicting high-resolution three-dimensional temperature and salinity field changes by using multi-source marine data, which comprises the following steps:

[0007] Step 1: Collect satellite remote sensing observed sea surface temperature, sea surface anomaly data and high-precision three-dimensional temperature and salinity data within the prediction time, and then construct sea surface temperature data set, sea surface anomaly data set and three-dimensional temperature and salinity data set; divide the three data sets according to the research area, construct training samples, and divide the training set, the validation set and the test set according to the time length; wherein the sea surface temperature training sample is constructed, including N continuous sampling time sea surface temperature remote sensing images, and the label is the K continuous sampling time sea surface temperature remote sensing image; the sea surface anomaly training sample is constructed, including N continuous sampling time sea surface anomaly remote sensing images, and the label is the K continuous sampling time sea surface anomaly remote sensing image; the three-dimensional temperature and salinity training sample is constructed, including the same sampling time sea surface temperature remote sensing image and sea surface anomaly remote sensing image, and the label is the same sampling time three-dimensional temperature and salinity data, and the same sampling time sea surface temperature remote sensing image and sea surface anomaly remote sensing image are interpolated to make the spatial resolution of the image the same as that of the three-dimensional temperature and salinity data.

[0008] Step 2: Construct sea surface temperature prediction model, sea surface anomaly prediction model and three-dimensional temperature and salinity prediction model. Select SimVP-gsta model, which is an encoder-spatio-temporal translator-decoder structure; the spatio-temporal translator in the SimVP-gsta model is realized by using Mobile-ViT module to construct M-ViT model; select M-ViT model to construct sea surface temperature prediction model and train using sea surface temperature training sample; select M-ViT model to construct sea surface anomaly prediction model and train using sea surface anomaly training sample; select SimVP-gsta model to construct three-dimensional temperature and salinity prediction model and train using three-dimensional temperature and salinity training sample.

[0009] Step 3: Obtain the optimal sea surface temperature prediction model, the sea level anomaly prediction model and the three-dimensional temperature and salinity prediction model under the sea after training, which are used to predict the change of the three-dimensional temperature and salinity field. Collect the sea surface temperature remote sensing images and the sea level anomaly remote sensing images at N continuous time points, and input the optimal sea surface temperature prediction model and the optimal sea level anomaly prediction model respectively, to predict the output of the sea surface temperature remote sensing images and the sea level anomaly remote sensing images at K continuous sampling time points; difference the sea surface temperature remote sensing images and the sea level anomaly remote sensing images at each sampling time point obtained by prediction, so that the spatial resolution of the images is the same as that of the three-dimensional temperature and salinity data under the sea, and then input the sea surface temperature remote sensing images and the sea level anomaly remote sensing images after difference at each sampling time point into the optimal three-dimensional temperature and salinity prediction model under the sea, to predict the output of the three-dimensional temperature and salinity data under the sea at each sampling time point.

[0010] Compared with the prior art, the advantages and beneficial effects of the present application include:

[0011] (1) The current three-dimensional temperature and salinity field prediction based on artificial intelligence relies on expensive three-dimensional data, and as the prediction length increases during training, the stacking of three-dimensional high-resolution data greatly increases the difficulty of model training. For example, the XiHe model is very slow when reading large amounts of three-dimensional data during training, and the high-resolution three-dimensional data information as the label data during training is stacked in terms of prediction efficiency, making training very difficult. The longer the prediction efficiency, the greater the burden on the computer to read data. For example, the XiHe model takes more than three months to train. The method of the present application uses satellite remote sensing data and high-resolution reanalysis data to train a deep learning model to construct a prediction model, uses satellite remote sensing data with low resolution as input, and different prior art uses expensive three-dimensional temperature and salinity data as input, greatly reducing the amount of data, solving the problem of stacking three-dimensional high-resolution data in the time dimension, reducing the difficulty of model training, saving a large amount of computing resources, reducing the prediction cost, and the constructed high-resolution three-dimensional temperature and salinity field prediction model is more practical.

[0012] (2) The method of the present application trains and learns a deep learning model based on historical satellite remote sensing sea surface temperature, sea level anomaly and temperature and salinity reanalysis data, converts three-dimensional prediction into a nonlinear fusion form of sea surface element prediction and undersea temperature and salinity inversion, and realizes accurate prediction of sea surface temperature and sea level anomaly through the M-ViT model. The data obtained by prediction is interpolated and then input into the SimVP-gsta model to predict the three-dimensional temperature and salinity data under the sea. Compared with the traditional numerical mode, the method of the present application saves a large amount of computing resources, and is far superior to the numerical mode in terms of prediction efficiency and calculation speed. Compared with the prior art, the method of the present application can comprehensively utilize multi-source ocean data, effectively increase the prediction efficiency without reducing the time and spatial resolution of the prediction result, and accurately predict the ocean temperature and salinity at most layers, with higher resource utilization efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 Flow chart of the method for predicting high-resolution three-dimensional temperature and salinity field changes using multi-source marine data according to the present application;

[0014] Figure 2 Data set division diagram for an embodiment of the present application; (a) is the construction of training samples and data set division for predicting satellite remote sensing sea surface temperature or sea surface anomaly data, and (b) is the construction of training samples and data set division for inverting subsurface three-dimensional temperature and salinity data;

[0015] Figure 3 Overall structure diagram of the SimVP-gsta model;

[0016] Figure 4 Structure diagram of the gsta translator in the SimVP-gsta model;

[0017] Figure 5 Overall structure diagram of the M-ViT model used in an embodiment of the present application;

[0018] Figure 6 Structure diagram of the Mobile-ViT translator used in the M-ViT model of an embodiment of the present application;

[0019] Figure 7 Structure diagram of the three-dimensional temperature and salinity field prediction model overall implemented in an embodiment of the present application. DETAILED DESCRIPTION

[0020] The present application will be further described in detail below with reference to the accompanying drawings and embodiments. Three-dimensional temperature and salinity prediction method

[0021] The method for predicting high-resolution three-dimensional temperature and salinity field changes using multi-source marine data provided by the present application relies on the nonlinear relationship between sea surface temperature and sea surface anomaly and subsurface three-dimensional temperature and salinity field, and converts three-dimensional temperature and salinity prediction into horizontal direction marine surface element prediction and vertical direction marine surface to subsurface temperature and salinity inversion. First, low-resolution satellite remote sensing sea surface temperature and sea surface anomaly data of 1 / 4° are used to predict sea surface temperature and sea surface anomaly, and then the prediction results are bilinearly interpolated and spatiotemporally matched with high-precision temperature and salinity reanalysis data to achieve high-resolution inversion of 1 / 12° of the three-dimensional temperature and salinity field under the sea. The method of the present application uses low-resolution satellite remote sensing sea surface temperature and sea surface anomaly data as input to achieve high-resolution inversion of the three-dimensional temperature and salinity field under the sea, rather than using three-dimensional temperature and salinity field data as input, thereby avoiding the stacking of three-dimensional high-resolution data, saving a large amount of computing resources, being more suitable for practical application, and being far superior to the prediction timeliness and computing speed of numerical models.

[0022] As Figure 1As shown, the method for predicting high-resolution three-dimensional temperature and salinity field changes by the embodiment of the application using multi-source marine data includes the following six steps.

[0023] Step 1, collect satellite remote sensing observed sea surface temperature, sea level anomaly data and high-precision three-dimensional temperature and salinity data under the sea.

[0024] The embodiment of the application collects global sea surface temperature (SST) data from the NOAA website, and the spatial resolution of the satellite remote sensing SST data is 1 / 4°, and the time resolution is daily average; global satellite remote sensing sea level anomaly (SLA) data and global high-precision marine reanalysis data are collected from the CMEMS website, the spatial resolution of the satellite remote sensing SLA data is 1 / 4°, and the time resolution is daily average, and the SLA is the difference between the measured sea surface height and the average sea surface height; the high-precision reanalysis data has 50 layers, the spatial resolution is 1 / 12°, and the time resolution is daily average, and the embodiment of the application needs to obtain the high-precision reanalysis three-dimensional temperature and salinity data of the remaining 49 layers except the sea surface. The embodiment of the application selects the three kinds of data in the period of 2006-2022.

[0025] Step 2, divide the obtained three data sets into two data sets of China offshore and northwest Pacific Ocean, and divide the data sets into training set, validation set and test set, and normalize the data.

[0026] The embodiment of the application divides the research area into China offshore (0-42°N, 105°E-130°E) and northwest Pacific Ocean (10°S-42°N, 99°E-180°E), and constructs two regional data sets. The 12-year data from 2006 to 2017 is used as the training set, the two-year data from 2018 to 2020 is used as the validation set, and the data from December 2021 to November 2022 is used as the test set. In order to accelerate the convergence of the model, the input data is subjected to Z-score normalization processing before being input into the model.

[0027] As Figure 2As shown, (a) is the training sample construction and dataset division of satellite remote sensing sea surface temperature or sea level anomaly data, wherein each training sample is a continuous 30-day data, and the label is the data of the last 20 days, that is, by inputting 30-day historical data into the model, the future 20-day data is predicted, and the sample is constructed by using the window sliding method. (b) is the sample construction and dataset division of the inverted three-dimensional temperature and salinity data under the sea, wherein each training sample is a one-day sea surface temperature grid map and a sea level anomaly grid map, which is a grid map with a spatial resolution of 1 / 4°. The label is a one-day three-dimensional temperature and salinity map of 49 layers except the sea surface with a spatial resolution of 1 / 12°, that is, the three-dimensional temperature and salinity data under the sea surface is predicted by combining the sea surface temperature and the sea level anomaly.

[0028] Step 3, constructing a sea surface temperature prediction model and a sea level anomaly prediction model for predicting low-resolution satellite remote sensing sea surface temperature and sea level anomaly, and constructing a three-dimensional temperature and salinity prediction model under the sea for predicting high-resolution three-dimensional temperature and salinity data under the sea.

[0029] In the embodiment of the application, the input of the sea surface temperature prediction model is the historical 30-day sea surface temperature remote sensing image data in the above-mentioned training sample, and the output is the future 20-day sea surface temperature remote sensing image data. The spatial resolution of the sea surface temperature of the remote sensing image is 1 / 4°. Similarly, the input of the sea level anomaly prediction model is the historical 30-day sea level anomaly remote sensing image data in the above-mentioned training sample, and the output is the future 20-day sea level anomaly remote sensing image data. The spatial resolution of the sea level anomaly data of the remote sensing image is 1 / 4°.

[0030] The input of the three-dimensional temperature and salinity prediction model under the sea of the application is the interpolated one-day sea surface temperature image and sea level anomaly image in the above-mentioned training sample, and the output is the three-dimensional temperature and salinity map of 49 layers under the sea on the same day. After obtaining the remote sensing images of the sea surface temperature and the sea level anomaly, the high-precision three-dimensional temperature and salinity data under the sea are first spatiotemporally matched, that is, the interpolation operation is performed, and then the matched sea surface temperature data and sea level anomaly data are combined as input data to invert the three-dimensional temperature and salinity data under the sea.

[0031] Due to the characteristics of large amount of information, multi-source data and large number of grids of the data input into the prediction model, the complex information carried in the data is not conducive to the learning of the model. How to make the model learn a simpler data structure and obtain a universal law is the key to building the model. The encoding of the UNet model can reduce the data feature information and reduce the computational complexity to improve the learning efficiency of the model, and the decoding can convert the low-resolution feature map into a high-resolution output to capture image details and structures and fuse with the context information in the decoding stage. The SimVP-gsta model adds a decoding structure, a spatiotemporal attention mechanism gsta module based on the original encoding-decoding structure, which can accelerate the convergence speed of the model and improve the running efficiency.

[0032] The embodiment of the present application is based on the SimVP-gsta model constructed in document 4 (Cheng Tan, Zhangyang Gao, Siyuan Li, and Stan Z. Li. 2022. SimVP: Towards simple yet powerful spatiotemporal predictive learning. arXiv preprint arXiv:2211.12509 (2022).), as shown in Figure 3 The SimVP-gsta model as a whole obeys the encoder-spatiotemporal translator-decoder architecture. In the embodiment of the present application, the encoder is composed of Ns encoding units, each of which is composed of two convolution layers stacked together, the convolution kernels of the two convolution layers are the same, both are 3*3, and the step lengths are 1 and 2 respectively, each convolution layer is composed of 1 two-dimensional convolution (Conv2D) and 1 group normalization (Groupnorm) and an activation function SiLU; the decoder is also composed of Ns decoding units, each of which is composed of a deconvolution layer and a convolution layer stacked together, the convolution kernels are the same, both are 3*3, and the step lengths are both 1, each deconvolution layer / convolution layer is composed of 1 Conv2D, up-sampling PixelShuffle, 1 Groupnorm and an activation function SiLU; the spatiotemporal translator is located between the encoder and the decoder. The spatiotemporal translator is composed of Nt gsta modules, the structure of each gsta module is as shown in Figure 4 .

[0033] The embodiment of the present application proposes an improved M-ViT model based on the above-mentioned SimVP-gsta model. As shown in Figure 5 The structure of the M-ViT model is similar to that of the SimVP-gsta model, and the overall architecture obeys the encoder-spatiotemporal translator-decoder architecture, and the structures of the encoder and the decoder are similar to those of Figure 3The same as in the SimVP-gsta model, but the translator uses a lightweight Mobile-ViT module. Figure 6 As shown in the figure, the convolution structure of the Mobile-ViT model realizes local attention, and the self-attention structure realizes global attention, which has good spatiotemporal conversion performance for high-dimensional spatiotemporal data.

[0034] Given an input data, the data dimension is (B, X, H, W), where B is the input batch, X is the spatiotemporal layer, H is the height of the data, and W is the width of the data. The height and width of the data can be determined according to the spatial resolution and the size of the study area. The data is input into the M-ViT model, and the encoder will downsample the input data to reduce the complexity of the data in time and space. The dimension of the data obtained after the encoder processing is (B, X1, H1, W1). The encoding result is input into the spatiotemporal translator, which learns the changes between frames and learns the spatiotemporal evolution law. Then the translation result is input into the decoder, which will upsample the input data and splice the output result of the first encoding unit, and then pass through the last decoding unit to obtain the final output result.

[0035] The embodiments of the present application are aimed at three problems of sea surface temperature prediction, sea level anomaly prediction and three-dimensional temperature and salinity prediction under water. For each prediction problem, an M-ViT model, a SimVP-gsta model and a UNet model are used to construct a prediction model, and then the data set obtained in step 2 is used to compare and train the three prediction models under the three prediction problems, and the optimal prediction model under each prediction problem is selected.

[0036] Step 4, train the prediction model, and constantly adjust the learning rate and model parameters to obtain the optimal prediction model.

[0037] In the training process, the loss function used by the embodiments of the present application is the function, and the formula is as follows:

[0038]

[0039] Where N is the total number of grids of the prediction model output data, p i is the predicted data output by the model, y i is the label data, and i represents the i-th grid point. The loss function is used for back propagation to update the gradient, and the model is optimized to make the predicted value approach the true value.

[0040] The Adam optimizer is used to train the model, and the learning rate adjustment strategy is adjusted to OneCycleLR. The channel number, learning rate and Batch_size of the model are constantly adjusted, and the optimal model is saved.

[0041] The sea surface temperature and sea level anomaly data are bilinearly interpolated and reanalyzed temperature and salinity data are spatially matched, the matched sea surface temperature data and sea level anomaly data are combined as input data, the reanalyzed data are taken as label data, the M-ViT model, the SimVP-gsta model and the UNet model are respectively compared and trained, and the three models are trained to be optimal.

[0042] Step 5, for the three M-ViT models, SimVP-gsta models and UNet models obtained under each prediction problem, the optimal prediction model is selected from the comparison.

[0043] The test data set is input into the trained model by seasons, December, 2021-February, 2022 is regarded as winter, March, 2022-May, 2022 is regarded as spring, June, 2022-August, 2022 is regarded as summer, and September, 2022-November, 2022 is regarded as autumn. The prediction effect is evaluated by using MSE, MAE and RMSE, and the prediction ability of the model in each season is observed and counted. The input of the model each time is 30 days of historical satellite remote sensing data, and the three-dimensional temperature and salinity data in the future 20 days are predicted.

[0044] After the test experiment, the sea surface temperature prediction model and the sea level anomaly prediction model constructed by the M-ViT model are optimal, and the three-dimensional temperature and salinity prediction model constructed by the SimVP-gsta model is optimal. Therefore, the method of the application will use 1 / 4° low-resolution satellite remote sensing data to predict the sea surface temperature and sea level anomaly by the M-ViT model, then match the prediction results with the reanalyzed data in time and space, and then use the SimVP-gsta model to invert the three-dimensional temperature and salinity field under the sea, so as to finally realize the 1 / 12° high-resolution three-dimensional temperature and salinity prediction model.

[0045] Step 6, the trained low-resolution prediction model is combined with the high-resolution inversion model to obtain the final high-resolution three-dimensional temperature and salinity prediction model.

[0046] As Figure 7As shown, the three-dimensional temperature and salinity prediction method of the method of the present application relies on the nonlinear relationship between sea surface temperature and sea level anomaly and subsurface three-dimensional temperature and salinity field, and converts three-dimensional temperature and salinity prediction into horizontal direction ocean surface element prediction and vertical direction ocean surface to subsurface temperature and salinity inversion. After obtaining the optimal sea surface temperature prediction model, the sea level anomaly prediction model and the subsurface three-dimensional temperature and salinity prediction model, the steps of the embodiment of the present application for high-resolution three-dimensional temperature and salinity prediction are: first, obtaining 1 / 4° low-resolution satellite remote sensing sea surface temperature data and sea level anomaly data, using the optimal sea surface temperature prediction model and the optimal sea level anomaly prediction model to predict the sea surface temperature and the sea level anomaly, respectively, then performing spatiotemporal matching of the prediction results of the sea surface temperature and the sea level anomaly with the reanalysis data, and then using the SimVP-gsta model to invert the subsurface three-dimensional temperature and salinity field, thereby realizing a 1 / 12° high-resolution three-dimensional temperature and salinity prediction model.

[0047] Since the input data of the embodiment of the present application is subjected to Z-score normalization processing before being input into the prediction model, the output prediction data is subjected to inverse Z-score normalization processing, and the final high-resolution subsurface 49-layer three-dimensional temperature and salinity prediction data is obtained.

[0048] Currently, three-dimensional temperature and salinity field prediction is mostly based on numerical mode prediction algorithm, but it has the disadvantages of short prediction time, slow calculation speed and large calculation resource consumption. Deep learning algorithm can well overcome these shortcomings compared with numerical mode, but at present, three-dimensional temperature and salinity field prediction based on deep learning algorithm is mostly based on expensive three-dimensional data, and with the increase of prediction time, three-dimensional data stacking will greatly increase the model burden, so the current research mostly has the problem of insufficient prediction time or spatial resolution. As can be seen from the above description, the method of the present application uses deep learning algorithm to establish the nonlinear relationship between the sea surface data at the historical time and the subsurface temperature and salinity field at the future time, converts three-dimensional temperature and salinity prediction into sea surface element prediction and sea surface element joint inversion of subsurface temperature and salinity, uses 1 / 4° low-resolution satellite remote sensing data to construct the prediction model of sea surface temperature and sea level anomaly, and realizes the construction of the inversion model of SST&SLA joint subsurface temperature and salinity by using 1 / 12° high-resolution temperature and salinity data as labels, and integrates the prediction model and the inversion model to obtain the final high-resolution three-dimensional temperature and salinity prediction model. The method of the present application can well solve the shortcomings of insufficient prediction time or spatial resolution of current deep learning, avoid the problem of large amount of data caused by stacking of high-resolution historical data, and can effectively predict the future temperature and salinity field change.

[0049] All technical features described in the specification are known to the person skilled in the art. The present application omits the description of the known components and known technologies to avoid redundancy and unnecessarily limit the present application. The embodiments described in the above examples also do not represent all embodiments consistent with the present application. Various modifications or variations made by the person skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A method for predicting high-resolution three-dimensional temperature and salinity field changes using multi-source ocean data, characterized in that: Comprising the following steps: Step 1: Collect satellite remote sensing observation of sea surface temperature, sea level anomaly data and high-precision three-dimensional temperature and salinity data under the sea within the predicted time, and then build three data sets: sea surface temperature data set, sea level anomaly data set and three-dimensional temperature and salinity data set under the sea; The three data sets are respectively divided according to the research area, the training samples are constructed, and the training set, the verification set and the test set are divided according to the time length; Among them, the sea surface temperature training sample is constructed, including N continuous sampling time sea surface temperature remote sensing image, the label is the latter K continuous sampling time sea surface temperature remote sensing image; The sea level anomaly training sample is constructed, including N continuous sampling time sea level anomaly remote sensing image, the label is the latter K continuous sampling time sea level anomaly remote sensing image; The three-dimensional temperature and salinity training sample under the sea is constructed, including the same sampling time sea surface temperature remote sensing image and sea level anomaly remote sensing image, the label is the same sampling time three-dimensional temperature and salinity data under the sea, and the same sampling time sea surface temperature remote sensing image and sea level anomaly remote sensing image are interpolated, so that the spatial resolution of the image is the same as that of the three-dimensional temperature and salinity data under the sea; Step 2: Construct sea surface temperature prediction model, sea level anomaly prediction model and three-dimensional temperature and salinity prediction model under the sea; Select SimVP-gsta model, which is an encoder-spatiotemporal translator-decoder structure; The spatiotemporal translator in the SimVP-gsta model is realized by using Mobile-ViT module to construct M-ViT model; Select M-ViT model to construct sea surface temperature prediction model, and train it using sea surface temperature training sample; Select M-ViT model to construct sea level anomaly prediction model, and train it using sea level anomaly training sample; Select SimVP-gsta model to construct three-dimensional temperature and salinity prediction model under the sea, and train it using three-dimensional temperature and salinity training sample under the sea; Step 3: Obtain the optimal sea surface temperature prediction model, sea level anomaly prediction model and three-dimensional temperature and salinity prediction model under the sea after training, which are used to predict the change of three-dimensional temperature and salinity field; Collect sea surface temperature remote sensing image and sea level anomaly remote sensing image at continuous N time, input the optimal sea surface temperature prediction model and the optimal sea level anomaly prediction model respectively, and predict the output of the latter K continuous sampling time sea surface temperature remote sensing image and sea level anomaly remote sensing image; The sea surface temperature remote sensing image and sea level anomaly remote sensing image obtained by prediction at each sampling time are interpolated, so that the spatial resolution of the image is the same as that of the three-dimensional temperature and salinity data under the sea, and then each sampling time sea surface temperature remote sensing image and sea level anomaly remote sensing image after interpolation are input into the optimal three-dimensional temperature and salinity prediction model under the sea, and the three-dimensional temperature and salinity data under the sea at each sampling time are predicted and output.

2. The method of claim 1, wherein, In step 1, the global sea surface temperature satellite remote sensing image data is collected, and the spatial resolution of the image is 1 / 4° and the time resolution is daily; Global sea level anomaly satellite remote sensing image data is collected, and the spatial resolution of the image is 1 / 4° and the time resolution is daily; The global high-precision marine reanalysis data is collected to obtain three-dimensional temperature and salinity image data of 49 layers under the sea, and the spatial resolution of the image is 1 / 12° and the time resolution is daily.

3. The method according to claim 1 or 2, characterized in that, In step 1, the samples in the three data sets are subjected to Z-score normalization processing.

4. The method according to claim 1 or 2, characterized in that, In step 2, the M-ViT model is selected as an encoder-spatiotemporal translator-decoder structure; the encoder is composed of Ns encoding units, each of which is composed of two convolution layers stacked together, and the convolution kernels of the two convolution layers are the same, both being 3*3, and the steps are 1 and 2 in turn; The decoder is composed of Ns decoding units, each of which is composed of an inverse convolution layer and a convolution layer stacked together, and the convolution kernels are the same, both being 3*3, and the steps are both 1; the spatiotemporal translator is realized by using a Mobile-ViT module.

5. The method according to claim 1 or 2, characterized in that, In step 2, for the prediction of sea surface temperature, sea level anomaly and three-dimensional temperature and salinity under the sea, an M-ViT model, a SimVP-gsta model and a UNet model are used to construct a prediction model, and the training sample set obtained in step 1 is used to train the prediction model, and the optimal prediction model is selected by comparison.

Images