Method for predicting change of high-resolution three-dimensional temperature and salt field by using multi-source ocean data

Abstract

The invention discloses a method for predicting the change of a high-resolution three-dimensional temperature and salt field by using multi-source ocean data. According to the method, three-dimensional temperature and salinity forecasting is converted into ocean surface element forecasting in the horizontal direction and joint inversion of subsea surface temperature and salinity, firstly, a sea surface temperature forecasting model and a sea level anomaly forecasting model are constructed by using low-resolution satellite remote sensing data and using an M-ViT model; carrying out difference value matching on the predicted sea surface temperature and sea level anomaly prediction data and high-resolution undersea three-dimensional thermohaline data, and constructing an undersea three-dimensional thermohaline prediction model by using a SimVP-gsta model to realize undersea three-dimensional thermohaline high-resolution prediction on the sea surface temperature and sea level anomaly low-resolution data. According to the method, the data size problem caused by stacking of high-resolution historical data is avoided, the prediction time efficiency can be effectively improved on the premise that the time resolution and the space resolution of a prediction result are not reduced, and the ocean temperature and salinity can be accurately predicted in most layers.

Description

Technical Field

[0001] The invention belongs to the technical field of ocean temperature-salinity field prediction, and in particular relates to a method for predicting high-resolution three-dimensional temperature-salinity field changes by using multi-source ocean data. Background Art

[0002] High-resolution three-dimensional temperature and salinity data can provide refined and comprehensive data support and analysis for marine environmental monitoring. These data not only help to more accurately 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. By monitoring and analyzing high-resolution three-dimensional temperature and salinity fields, scientists can more accurately predict changes in ocean temperature and salinity fields, thereby improving their ability to respond to extreme weather events, marine disaster warnings, and marine environmental protection. Therefore, improving the accuracy of ocean three-dimensional temperature and salinity field predictions has far-reaching significance for promoting marine scientific research and promoting marine environmental protection.

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

[0004] At present, with the rapid development of data-driven deep learning methods, ocean forecasting methods based on deep learning are considered to be a powerful supplement to traditional methods. Related models of spatiotemporal series prediction based on deep learning are applied to the spatiotemporal prediction of sea temperature. These models can capture the spatiotemporal correlation and nonlinear characteristics in sea temperature data. By learning from historical meteorological data, they can discover the hidden patterns and laws in the climate system, and then predict the trend of future sea temperature changes. However, in the past, most studies on ocean temperature and salinity prediction focused on the two-dimensional plane of the sea surface. However, for marine environmental monitoring, effective prediction in three-dimensional space is more worthy of attention. In recent years, Zuo et al. used deep learning models to predict ocean temperature, but their prediction time is only one day, and the prediction area is small, which is difficult to meet the needs of marine 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 the 3D-UNet model to predict the seafloor temperature (SSbT) above 400 meters in the Pacific Ocean and its adjacent oceans, but the temporal resolution of the data was monthly average data and the spatial resolution was only 1°×1° (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 the prediction time, it took more than three months to train on 8 NAVIDAH100s (Reference 3: WangX, Wang R, Hu N, et al. Xihe: A data-driven model for global ocean eddy-resolvingforecasting [J]. arXiv preprint arXiv: 2402.02995, 2024.). Summary of the invention

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

[0006] The present invention provides a method for predicting high-resolution three-dimensional temperature-salinity field changes using multi-source ocean data, comprising the following steps:

[0007] Step 1: Collect the sea surface temperature, sea level anomaly data and high-precision three-dimensional underwater temperature and salinity data observed by satellite remote sensing within the prediction time, and then construct the sea surface temperature dataset, sea level anomaly dataset and underwater three-dimensional temperature and salinity dataset; divide the three datasets according to the study area, construct training samples, and divide the training set, validation set and test set according to the time length; among them, construct the sea surface temperature training samples, including the sea surface temperature remote sensing images at N consecutive sampling moments, and the labels are the sea surface temperature remote sensing images at the next K consecutive sampling moments; construct the sea level anomaly training samples, including the sea level anomaly remote sensing images at N consecutive sampling moments, and the labels are the sea level anomaly remote sensing images at the next K consecutive sampling moments; construct the underwater three-dimensional temperature and salinity training samples, including the sea surface temperature remote sensing images and sea level anomaly remote sensing images at the same sampling moment, and the labels are the underwater three-dimensional temperature and salinity data at the same sampling moment, and interpolate the sea surface temperature remote sensing images and sea level anomaly remote sensing images at the same sampling moment so that the spatial resolution of the images is the same as that of the underwater three-dimensional temperature and salinity data.

[0008] Step 2: Construct sea surface temperature prediction model, sea level anomaly prediction model and underwater three-dimensional temperature and salinity prediction model. Select the SimVP-gsta model, which is an encoder-space-time translator-decoder structure; construct the M-ViT model by implementing the space-time translator in the SimVP-gsta model using the Mobile-ViT module; select the M-ViT model to construct the sea surface temperature prediction model, and use the sea surface temperature training samples for training; select the M-ViT model to construct the sea level anomaly prediction model, and use the sea level anomaly training samples for training; select the SimVP-gsta model to construct the underwater three-dimensional temperature and salinity prediction model, and use the underwater three-dimensional temperature and salinity training samples for training.

[0009] Step 3: Obtain the optimal sea surface temperature prediction model, sea level anomaly prediction model and underwater three-dimensional temperature and salinity prediction model after training, which are used to predict the changes in the three-dimensional temperature and salinity field. Collect sea surface temperature remote sensing images and sea level anomaly remote sensing images at N consecutive moments, input them into the optimal sea surface temperature prediction model and the optimal sea level anomaly prediction model respectively, and predict and output the sea surface temperature remote sensing images and sea level anomaly remote sensing images at the next K consecutive sampling moments; perform difference calculation on the predicted sea surface temperature remote sensing images and sea level anomaly remote sensing images at each sampling moment, so that the spatial resolution of the images is the same as the spatial resolution of the underwater three-dimensional temperature and salinity data, and then input the differenced sea surface temperature remote sensing images and sea level anomaly remote sensing images at each sampling moment into the optimal underwater three-dimensional temperature and salinity prediction model, and predict and output the underwater three-dimensional temperature and salinity data at each sampling moment.

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

[0011] (1) At present, the three-dimensional temperature-salinity field prediction based on artificial intelligence relies on expensive three-dimensional data, and as the prediction time increases during training, the stacking of three-dimensional high-resolution data will greatly increase the difficulty of model training. For example, the XiHe model is very slow when reading a large amount of three-dimensional data during training. At the same time, the high-resolution three-dimensional data information is stacked as label data during training in the prediction time, which makes training very difficult. The longer the prediction time, the greater the burden on the computer to read data. For example, the training of the XiHe model takes more than 3 months. The method of the present invention uses satellite remote sensing data and high-resolution reanalysis data to train a deep learning model to construct a prediction model, and uses low-resolution satellite remote sensing data of the sea surface as input. Unlike the existing technology, which uses expensive three-dimensional temperature-salinity data as input, it greatly reduces the amount of data, solves the current problem of stacking three-dimensional high-resolution data in the time dimension, reduces the difficulty of model training, saves a lot of computing resources, and reduces the prediction cost. The constructed high-resolution three-dimensional temperature-salinity field prediction model is more suitable for practical applications.

[0012] (2) The method of the present invention trains and learns the deep learning model based on historical satellite remote sensing sea surface temperature, sea level anomaly and temperature-salinity reanalysis data, converts the three-dimensional forecast into a nonlinear fusion form of sea surface element prediction and underwater temperature-salinity inversion, and can achieve accurate prediction of sea surface temperature and sea level anomaly through the M-ViT model. The predicted data is interpolated and then predicted through the SimVP-gsta model to obtain underwater three-dimensional temperature-salinity data. Compared with traditional numerical models, the method of the present invention saves a lot of computing resources, and is far superior to numerical models in terms of forecast timeliness and calculation speed. Compared with the prior art, the method of the present invention can comprehensively utilize multi-source ocean data, effectively increase the forecast timeliness without reducing the temporal and spatial resolution of the forecast results, and can accurately predict ocean temperature and salinity in most layers, with higher resource utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A flow chart of the method for predicting high-resolution three-dimensional temperature-salinity field changes using multi-source ocean data of the present invention;

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

[0015] Figure 3 It is the overall structure diagram of the SimVP-gsta model;

[0016] Figure 4 This is the structure diagram of the gsta translator in the SimVP-gsta model;

[0017] Figure 5 The overall structure diagram of the M-ViT model used in the embodiment of the present invention;

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

[0019] Figure 7 This is a structural diagram of the three-dimensional temperature-salinity field prediction model implemented as a whole in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0021] The method for predicting high-resolution three-dimensional temperature-salinity field changes using multi-source ocean data provided by the present invention relies on the nonlinear relationship between sea surface temperature and sea level anomaly and sub-surface three-dimensional temperature-salinity field, and converts the three-dimensional temperature-salinity forecast into the forecast of ocean surface elements in the horizontal direction and the inversion of sub-surface temperature-salinity in the vertical direction of the ocean surface. First, the sea surface temperature and sea level anomaly data of 1 / 4° low-resolution satellite remote sensing are used to forecast the sea surface temperature and sea level anomaly, and then the forecast results are bilinearly interpolated and matched with high-precision temperature-salinity reanalysis data in time and space to achieve 1 / 12° high-resolution inversion of the three-dimensional temperature-salinity field under the sea. The method of the present invention uses low-resolution satellite remote sensing sea surface temperature and sea level anomaly data as input to achieve high-resolution inversion of the three-dimensional temperature-salinity field under the sea, rather than using three-dimensional temperature-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 applications, and being far superior to the forecast timeliness of numerical models in terms of forecast timeliness and calculation speed.

[0022] like Figure 1As shown, the method for predicting high-resolution three-dimensional temperature-salinity field changes using multi-source ocean data in an embodiment of the present invention includes the following six steps.

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

[0024] The embodiment of the present invention collects global sea surface temperature (SST) data from the NOAA official website. The spatial resolution of satellite remote sensing SST data is 1 / 4°, and the temporal resolution is daily average; collects global satellite remote sensing sea level anomaly (SLA) data and global high-precision ocean reanalysis data from the CMEMS official website. The spatial resolution of satellite remote sensing SLA data is 1 / 4°, and the temporal resolution is daily average. SLA is the difference between the measured sea surface height and the average sea surface height; the high-precision reanalysis data has a total of 50 layers, with a spatial resolution of 1 / 12° and a temporal resolution of daily average. The embodiment of the present invention needs to obtain the remaining 49 layers of high-precision reanalysis subsurface three-dimensional temperature and salinity data except the sea surface. The embodiment of the present invention chooses to collect these three types of data for the period of 2006-2022.

[0025] Step 2: Divide the three data sets obtained into two regions, namely the China coastal waters and the northwest Pacific Ocean, to construct two data sets, divide the data sets into training sets, validation sets, and test sets, and normalize the data.

[0026] The embodiment of the present invention divides the study area into two parts: the offshore waters of China (0-42°N, 105°E-130°E) and the northwest Pacific Ocean (10°S-42°N, 99°E-180°E), and constructs data sets for the two regions accordingly. 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 model convergence, the input data is Z-score normalized before being input into the model.

[0027] like Figure 2As shown in the figure, (a) is the training sample construction and data set division of satellite remote sensing sea surface temperature or sea level anomaly data, where each training sample is a continuous 30-day data, and the label is the data of the next 20 days, that is, by inputting 30 days of historical data into the model, the data of the next 20 days is predicted, and the sample is constructed by the window sliding method. (b) is the sample construction and data set division of the inverted underwater three-dimensional temperature and salinity data, where each training sample is a day's sea surface temperature grid map and sea level anomaly grid map, which is a grid map with a spatial resolution of 1 / 4° interpolated to obtain a grid map with a spatial resolution of 1 / 12°, and the label is a day's three-dimensional temperature and salinity map of 49 layers except the sea surface with a spatial resolution of 1 / 12°, that is, the sub-surface three-dimensional temperature and salinity data are predicted by combining sea surface temperature and sea level anomaly.

[0028] Step 3: construct a sea surface temperature prediction model and a sea level anomaly prediction model for the prediction of low-resolution satellite remote sensing sea surface temperature and sea level anomalies, and construct a three-dimensional underwater temperature and salinity prediction model for the prediction of high-resolution three-dimensional underwater temperature and salinity data.

[0029] In the embodiment of the present invention, the input of the sea surface temperature prediction model is the historical 30-day sea surface temperature remote sensing image data in the above training samples, and the output is the sea surface temperature remote sensing image data for the next 20 days. The spatial resolution of the sea surface temperature on 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 training samples, and the output is the sea level anomaly remote sensing image data for the next 20 days. The spatial resolution of the sea level anomaly data on the remote sensing image is 1 / 4°.

[0030] The input of the underwater three-dimensional temperature-salinity prediction model of the present invention is the interpolated sea surface temperature image and sea level anomaly image of one day in the above training samples, and the output is the three-dimensional temperature-salinity map of 49 layers under the sea on that day. After obtaining the remote sensing images of sea surface temperature and sea level anomaly, they are firstly matched with the high-precision underwater three-dimensional temperature-salinity data in time and space, that is, 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 underwater three-dimensional temperature-salinity data.

[0031] Because the data of the input prediction model has the characteristics of large amount of information, multi-source data, and large number of grids, 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 universal laws is the key to the model to be constructed in the present invention. The encoding of the UNet model can reduce data feature information and reduce computational complexity to improve the learning efficiency of the model, while decoding is performed by converting low-resolution feature maps into high-resolution outputs to capture image details and structures, and merged with the contextual information of the decoding stage. The SimVP-gsta model, based on the original encoding-decoding structure, adds a decoding structure-space-time attention mechanism gsta module, which can accelerate the convergence speed of the model and improve operating efficiency.

[0032] The embodiment of the present invention constructs the SimVP-gsta model based on 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).), such as Figure 3 As shown, the SimVP-gsta model as a whole complies with the encoder-space-time translator-decoder architecture. In an embodiment of the present invention, the encoder is composed of Ns encoding units, each encoding unit is composed of two superimposed convolutional layers, the convolution kernels of the two convolutional layers are the same, both are 3*3, the step sizes are 1 and 2 respectively, and each convolutional layer is composed of a 2D convolution (Conv2D) and a group normalization (Groupnorm) and an activation function SiLU; the decoder is also composed of Ns decoding units, each decoding unit is composed of a deconvolution layer and a convolution layer superimposed, the convolution kernels are the same, both are 3*3, the step size is 1, and each deconvolution layer / convolution layer is composed of a Conv2D, an upsampling PixelShuffle (pixel reorganization), a Groupnorm and an activation function SiLU; the space-time translator is located between the encoder and the decoder. The space-time translator is composed of Nt gsta modules, and the structure of each gsta module is as follows Figure 4 shown.

[0033] The embodiment of the present invention proposes an improved M-ViT model based on the above-mentioned SimVP-gsta model. Figure 5 As shown in Figure 2, the M-ViT model structure is similar to the SimVP-gsta model. The overall architecture follows the encoder-spatial-temporal translator-decoder architecture. The structures of its encoder and decoder are similar to Figure 3The translator is the same as in the SimVP-gsta model, but uses a lightweight Mobile-ViT module. The structure of the Mobile-ViT module is as follows 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] Suppose an input data is given, 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 research 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. Suppose the dimension of the data obtained after the encoder processing is (B, X1, H1, W1). The encoding result is input into the spatiotemporal translator, and the spatiotemporal translator learns the changes between each frame and the laws of spatiotemporal evolution. Then the translation result enters the decoder, and the decoder will upsample the input data and splice it with the output result of the first encoding unit, and then pass it through the last layer of decoding units to obtain the final output result.

[0035] The embodiment of the present invention aims at three problems: sea surface temperature prediction, sea level anomaly prediction and underwater three-dimensional temperature and salinity prediction. In each prediction problem, the M-ViT model, SimVP-gsta model and UNet model are used to build a prediction model. Then, the data set obtained in step 2 is used to perform comparative training on the three prediction models for the three prediction problems, and the optimal prediction model for each prediction problem is selected.

[0036] Step 4: Train the prediction model, continuously adjust the learning rate and model parameters, and obtain the optimal prediction model.

[0037] The embodiment of the present invention uses a loss function as a function during the training process, and the formula is as follows:

[0038]

[0039] Where N is the total number of grids for the output data of the prediction model, 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 gradient is updated through the back propagation of the loss function, and the model is optimized to make the predicted value close to the true value.

[0040] The Adam optimizer is used to train the model, and the learning rate adjustment strategy is adjusted to OneCycleLR. The number of channels, learning rate, and batch size of the model are continuously adjusted, and the optimal model is trained and saved.

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

[0042] Step 5: For each prediction problem, the three M-ViT models, SimVP-gsta model, and UNet model are compared and the best prediction model is selected.

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

[0044] After testing experiments, the present invention selects the sea surface temperature prediction model and sea level anomaly prediction model constructed by the M-ViT model as the best, and selects the underwater three-dimensional temperature and salinity prediction model constructed by the SimVP-gsta model as the best. Therefore, the method of the present invention will use 1 / 4° low-resolution satellite remote sensing data to predict sea surface temperature and sea level anomalies through the M-ViT model, then match the prediction results with the reanalysis data in time and space, and then use the SimVP-gsta model to invert the underwater three-dimensional temperature and salinity field, and finally realize a 1 / 12° high-resolution three-dimensional temperature and salinity prediction model.

[0045] Step 6: Use the trained low-resolution prediction model and the high-resolution inversion model to obtain the final high-resolution three-dimensional temperature-salinity prediction model.

[0046] like Figure 7As shown, the three-dimensional temperature-salinity forecasting method of the method of the present invention relies on the nonlinear relationship between the sea surface temperature and sea level anomaly and the subsurface three-dimensional temperature-salinity field, and converts the three-dimensional temperature-salinity forecast into the horizontal ocean surface element forecast and the vertical ocean surface to subsurface temperature-salinity inversion. After obtaining the optimal sea surface temperature prediction model, sea level anomaly prediction model and underwater three-dimensional temperature-salinity prediction model, the embodiment of the present invention performs high-resolution three-dimensional temperature-salinity prediction steps as follows: first, obtain 1 / 4° low-resolution satellite remote sensing sea surface temperature data and sea level anomaly data, use the optimal sea surface temperature prediction model and the optimal sea level anomaly prediction model to forecast sea surface temperature and sea level anomaly, then match the forecast results of sea surface temperature and sea level anomaly with reanalysis data in time and space, and then use the SimVP-gsta model to invert the underwater three-dimensional temperature-salinity field, thereby realizing a 1 / 12° high-resolution three-dimensional temperature-salinity forecasting model.

[0047] Since the input data of the embodiment of the present invention is subjected to Z-score normalization processing before being input into the prediction model, the output prediction data is further subjected to inverse Z-score normalization processing to obtain the final high-resolution three-dimensional temperature-salinity prediction data of 49 layers under the sea.

[0048] At present, the prediction of three-dimensional temperature and salinity is mostly based on numerical model prediction algorithms, but it has the disadvantages of short prediction time, slow calculation speed and large consumption of computing resources. Compared with numerical models, deep learning algorithms can overcome these shortcomings well, but the three-dimensional temperature and salinity prediction based on deep learning algorithms at this stage is mostly based on expensive three-dimensional data, and with the increase of prediction time, the stacking of three-dimensional data will greatly increase the model burden, so the current research has the problem of insufficient prediction time or spatial resolution. It can be seen from the above description that the method of the present invention uses a deep learning algorithm to establish a nonlinear relationship between the sea surface data at historical moments and the seabed temperature and salinity field at future moments, and converts the three-dimensional temperature and salinity prediction into the sea surface element prediction and the sea surface element joint inversion of subsurface temperature and salinity. First, the 1 / 4° low-resolution satellite remote sensing data is used to construct a prediction model for sea surface temperature and sea level anomalies, and then the 1 / 12° high-resolution temperature and salinity data is used as a label to realize the SST&SLA joint construction of the seabed temperature and salinity inversion model, and the prediction model and the inversion model are integrated to obtain the final high-resolution three-dimensional temperature and salinity prediction model. The method of the present invention can well solve the shortcomings of current deep learning prediction such as insufficient time or spatial resolution, avoid the data volume problem caused by stacking high-resolution historical data, and can effectively predict future changes in temperature and salinity fields.

[0049] Except for the technical features described in the specification, all of them are known technologies to those skilled in the art. The present invention omits the description of known components and known technologies to avoid redundancy and unnecessary limitation of the present invention. The implementation methods described in the above embodiments do not represent all implementation methods consistent with the present application. On the basis of the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative labor are still within the protection scope of the present invention.

Claims

1. A method for predicting high-resolution three-dimensional temperature-salinity field changes using multi-source ocean data. It is characterized in that The steps include: Step 1: Collect the sea surface temperature, sea level anomaly data and high-precision three-dimensional underwater temperature and salinity data observed by satellite remote sensing within the prediction time, and then construct three data sets: sea surface temperature data set, sea level anomaly data set and underwater 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, validation set and test set according to the time length; among them, construct the sea surface temperature training sample, including the sea surface temperature remote sensing images at N consecutive sampling moments, and the labels are the sea surface temperature remote sensing images at the next K consecutive sampling moments; construct the sea level anomaly training sample, including the sea level anomaly remote sensing images at N consecutive sampling moments, and the labels are the sea level anomaly remote sensing images at the next K consecutive sampling moments; construct the underwater three-dimensional temperature and salinity training sample, including the sea surface temperature remote sensing images and sea level anomaly remote sensing images at the same sampling moment, and the labels are the underwater three-dimensional temperature and salinity data at the same sampling moment, and interpolate the sea surface temperature remote sensing images and sea level anomaly remote sensing images at the same sampling moment, so that the spatial resolution of the image is the same as the spatial resolution of the underwater three-dimensional temperature and salinity data; Step 2: Construct sea surface temperature prediction model, sea level anomaly prediction model and underwater three-dimensional temperature and salinity prediction model; The SimVP-gsta model is selected, which is an encoder-space-time translator-decoder structure; the M-ViT model is constructed by implementing the space-time translator in the SimVP-gsta model using the Mobile-ViT module; The M-ViT model is selected to build a sea surface temperature prediction model, and the sea surface temperature training samples are used for training; The M-ViT model is selected to build a sea level anomaly prediction model, and the sea level anomaly training samples are used for training; The SimVP-gsta model was selected to construct a three-dimensional underwater temperature and salinity prediction model, and the three-dimensional underwater temperature and salinity training samples were used for training; Step 3: Obtain the optimal sea surface temperature prediction model, sea level anomaly prediction model and underwater three-dimensional temperature-salinity prediction model after training, which are used to predict the changes in the three-dimensional temperature-salinity field; The sea surface temperature remote sensing images and sea level anomaly remote sensing images of N consecutive moments are collected and respectively input into the optimal sea surface temperature prediction model and the optimal sea level anomaly prediction model, and the sea surface temperature remote sensing images and sea level anomaly remote sensing images of the next K consecutive sampling moments are predicted and output; the sea surface temperature remote sensing images and sea level anomaly remote sensing images of each sampling moment are interpolated to make the spatial resolution of the images the same as the spatial resolution of the underwater three-dimensional temperature and salinity data, and then the sea surface temperature remote sensing images and sea level anomaly remote sensing images after the interpolation at each sampling moment are input into the optimal underwater three-dimensional temperature and salinity prediction model, and the underwater three-dimensional temperature and salinity data at each sampling moment are predicted and output.

2. The method according to claim 1, It is characterized in that In the step 1, global sea surface temperature satellite remote sensing image data is collected, the spatial resolution of the image is 1 / 4°, and the temporal resolution is daily average; Collect global sea level anomaly satellite remote sensing image data, with a spatial resolution of 1 / 4° and a temporal resolution of daily average; Collect global high-precision ocean reanalysis data and obtain three-dimensional temperature and salinity image data of 49 layers under the sea. The spatial resolution of the image is 1 / 12° and the temporal resolution is daily average.

3. The method according to claim 1 or 2, It is characterized in that In step 1, Z-score normalization is performed on samples in the three data sets respectively.

4. The method according to claim 1 or 2, It is characterized in that In the step 2, the selected M-ViT model is an encoder-space-time translator-decoder structure; the encoder is composed of Ns coding units, each coding unit is composed of two superimposed convolutional layers, the convolution kernels of the two convolutional layers are the same, both are 3*3, and the step sizes are 1 and 2 respectively; The decoder consists 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, all 3*3, and the step size is 1. The spatiotemporal translator is implemented using the Mobile-ViT module.

5. The method according to claim 1 or 2, It is characterized in that In the step 2, a prediction model is constructed using the M-ViT model, SimVP-gsta model and UNet model for sea surface temperature prediction, sea level anomaly prediction and underwater three-dimensional temperature and salinity prediction, respectively. The prediction model is then trained using the training sample set obtained in step 1, and the optimal prediction model is selected by comparison.

Images