Remote sensing end-to-end ocean three-dimensional temperature forecasting method based on artificial intelligence

Abstract

The invention discloses a remote sensing end-to-end ocean three-dimensional temperature forecasting method based on artificial intelligence. The method comprises the following steps: acquiring multi-source remote sensing data and ocean reanalysis data, preprocessing the acquired multi-source remote sensing data and reanalysis data, constructing a space-time sequence forecasting model, training the space-time sequence forecasting model to be qualified, inputting the preprocessed multi-source remote sensing data into the space-time sequence forecasting model which is trained to be qualified, and forecasting the ocean reanalysis data according to the preprocessed space-time sequence forecasting model. And obtaining a forecast result, and carrying out visual display on the forecast result. According to the method, a remote sensing end-to-end forecasting mode is used, a generative space-time model architecture and a sliding window attention mechanism are provided, a space-time problem is disassembled into a space problem combined with a time dimension, and a global attention mechanism is simplified into a window attention mechanism by performing window segmentation in space, so that the calculation overhead is reduced, and the calculation efficiency is improved. And high-resolution ocean three-dimensional environment forecasting can be realized without inputting a three-dimensional numerical mode background field.

Description

Technical Field

[0001] This invention relates to the field of marine environmental specific variable measurement technology, specifically to an artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature forecasting method. Background Technology

[0002] The prediction of the three-dimensional structure of the ocean is of great significance for the study of ocean phenomena and the guidance of marine activities. Modern ocean forecasting mainly relies on statistical forecasting and numerical forecasting, with numerical forecasting based on dynamic numerical calculations and assimilation playing a dominant role. However, although numerical forecasting methods contribute very high forecast accuracy, the high computational cost of computational physics partial differential equations, the instability of nonlinear dynamic processes in the ocean, and the incomplete understanding of natural science laws by humankind limit further improvements in the accuracy of numerical forecasting. Currently, large-scale ocean models based on artificial intelligence have demonstrated superior forecast performance and computational speed; however, these models still rely on the three-dimensional background field generated by numerical simulation as input to produce forecast results, which limits their applicability. Furthermore, the need for higher temporal and spatial resolution leads to significant computational costs.

[0003] The development of satellite remote sensing has made it possible to study ocean changes with high resolution, large areas, and long durations. However, due to the limitations of the transmission medium, satellite remote sensing can only detect information about the ocean surface and cannot directly obtain observational data of the subsurface ocean. Currently, many scholars use artificial intelligence methods to invert subsurface ocean structure from satellite remote sensing images, but most of these are reconstructions based on historical data with low spatiotemporal resolution. There are relatively few studies that directly predict the three-dimensional structure of the ocean based on satellite remote sensing data as input. Most AI-based forecasting focuses on using the three-dimensional background field obtained from numerical models as input to predict the three-dimensional structure of the ocean, which has certain limitations.

[0004] Based on the above analysis, the problems and shortcomings of the existing technology are as follows: (1) Simple time or simple space modeling cannot capture the complex spatiotemporal characteristics of the ocean, while iterative spatiotemporal models cannot capture long-term spatiotemporal changes and are prone to error accumulation problems.

[0005] (2) At the level of artificial intelligence algorithms, the computational complexity of the global attention mechanism algorithm is quadratic. As the resolution of the computational data increases, the global attention mechanism will bring huge computational overhead.

[0006] (3) Ocean forecasting models based on the background field of three-dimensional numerical models as input still rely on the accuracy of numerical models, and the training cost is high. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing an artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature forecasting method.

[0008] To achieve the above objectives, this invention provides an artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature prediction method, comprising: Acquire multi-source remote sensing data and ocean reanalysis data, and preprocess the acquired multi-source remote sensing data and reanalysis data; A spatiotemporal sequence forecasting model is constructed, and the model is trained using historical ocean remote sensing data and reanalysis data, and the model is trained to a qualified state. The preprocessed multi-source remote sensing data is input into the trained spatiotemporal sequence prediction model to obtain prediction results, and the prediction results are then visualized.

[0009] Furthermore, the remote sensing data includes sea surface temperature, absolute dynamic height of sea surface, sea surface salinity, zonal wind speed, and meridional wind speed; the reanalysis dataset is a daily temperature dataset with a resolution of 1 / 12°, and its vertical spatial distribution is multi-layered.

[0010] Furthermore, the vertical spatial distribution of the reanalysis dataset consists of 23 layers, including 0.49m, 2.65m, 5.08m, 7.93m, 11.41m, 15.81m, 21.60m, 29.44m, 40.34m, 55.76m, 77.85m, 92.32m, 109.73m, 130.67m, 155.85m, 186.13m, 222.48m, 266.04m, 318.13m, 380.21m, 453.94m, 541.09m, and 643.57m.

[0011] Furthermore, the preprocessing method is as follows: Remote sensing images of different resolutions from multi-source remote sensing data are uniformly processed to a 1 / 12° resolution and a time resolution of 1 day. Replace missing data values ​​in multi-source remote sensing data with 0; Normalize multi-source remote sensing data; The remote sensing data and the reanalysis dataset were respectively integrated into spatiotemporally correlated tensors, with tensor dimensions of [missing information]. ,in, For batch quantity, For time step, For the number of channels, and It represents the length and width of the forecast sea area.

[0012] Furthermore, the spatiotemporal sequence prediction model includes: The first-dimensional reshaping module is used to reshape tensors. Shape reshaping ; A convolution module is used to reshape the tensor of the first dimension module through convolution. The encoding is done into a low-dimensional latent space to initially extract spatial feature information, as follows: ; in, This represents the initially extracted spatial feature information. This represents the activation function. This indicates a normalization operation. This represents the convolution operation. This represents the tensor output by the first-dimensional reshaping module; The channel attention module is used to emphasize or suppress information between channels and generate vectors containing key remote sensing features based on the spatial feature information initially extracted by the convolution module, as detailed below: ; in, This generates a vector containing key remote sensing feature information. Indicates channel attention. Indicates spatial attention, and Represents adaptive average pooling and max pooling. Represents the sigmoid function; The second-dimensional reshaping module is used to reshape the vector containing key remote sensing feature information generated by the channel attention module. ; Global position encoding module, used to in In terms of dimensions, the vectors containing key remote sensing features after shape reshaping by the second-dimensional reshaping module are encoded with 64 dimensions. After encoding, the spatiotemporal features are then compared with... The encoded values ​​are added together to generate a vector containing global location information, as shown below: ; in, It is a vector containing global position information. Indicates encoding operation; A multi-scale residual window spatiotemporally coupled attention module is used to extract spatiotemporal features from vectors containing global position information generated by the global position encoding module; The third-dimensional reshaping module is used to reshape the spatiotemporal features extracted by the multi-scale residual window spatiotemporally coupled attention module back to their original shape. ; The deconvolution module is used to decode the spatiotemporal features reshaped by the third-dimensional reshaping module and perform residual connection with the spatial feature information initially extracted by the convolution module to map the decoded spatiotemporal features onto the required profile depth, as follows: ; in, The mapping result output by the deconvolution module. This is a deconvolution operation; The fourth-dimensional reshaping module is used to reshape the mapping result of the deconvolution module to... Dimensions are used to obtain the forecast results.

[0013] Furthermore, the multi-scale residual window spatiotemporally coupled attention module includes: The window segmentation module is used to segment vectors containing global position information; The first attention mechanism module is used to extract the spatiotemporal features of the window from a vector containing global position information using the attention mechanism; The window downsampling module is used to resample the spatiotemporal features of the window in order to expand the receptive field of the attention mechanism. The second attention mechanism module is used to extract spatiotemporal features from the resampled data using the attention mechanism; The window upsampling module is used to restore the spatiotemporal features extracted by the second attention mechanism module to the window size, and then perform residual connection with the spatiotemporal features extracted by the first attention mechanism module to obtain spatiotemporal features of multiple windows. The window merging module is used to merge the spatiotemporal features of multiple windows to obtain the spatiotemporal features extracted by the multi-scale residual window spatiotemporally coupled attention module.

[0014] Furthermore, the spatiotemporal sequence prediction model is trained using the PyTorch framework with a learning rate of 0.001, an optimizer of Adam, a loss function of MSE, and an early stopping group of 30.

[0015] Furthermore, it also includes: performing error analysis on the forecast results of the spatiotemporal series forecasting model, wherein the evaluation index of the error analysis includes root mean square error. and coefficient of determination The specific calculation method is as follows: ; ; ; in, denoted as mean squared error, and m is the sample size of the forecast results. The actual value corresponding to the forecast result. The forecast results are from the spatiotemporal series forecasting model. This is the average of the true values.

[0016] Beneficial effects: Unlike iterative spatiotemporal sequence models and global spatiotemporal attention mechanisms, this invention uses a generative spatiotemporal model architecture and a sliding window attention mechanism. By performing window segmentation in space, the global attention mechanism is simplified into a window attention mechanism, which significantly reduces computational overhead and enables high-resolution three-dimensional marine environmental forecasting. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature forecasting method according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the spatiotemporal sequence prediction model according to an embodiment of the present invention; Figure 3 This is a comparison chart of the effects of the present invention and existing artificial intelligence methods. Detailed Implementation

[0018] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. These embodiments are implemented based on the technical solutions of the present invention, and it should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0019] like Figure 1 As shown, this embodiment of the invention provides an artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature prediction method, including: Multi-source remote sensing data and ocean reanalysis data were acquired and preprocessed. The remote sensing data included sea surface temperature (SST), sea surface absolute dynamic height (ADT), sea surface salinity (SSS), zonal wind speed (UWND), and meridional wind speed (VWND). SST data came from the global in-situ satellite observation dataset of The Operational Sea Surface Temperature and Ice Analysis (OSTIA) system, provided by the UK Met Office. ADT data came from the global sea surface height gridded L4 data product of the Copernicus Climate Change Service. SSS data came from the multi-source satellite sea surface salinity and density observation L4 data product provided by Consiglio Nazionale delle Ricerche (CNR), which integrates salinity data from SMOS and SMAP satellites. Sea surface wind field data included zonal wind speed (UWND) and meridional wind speed (VWND), provided by the global gridded wind field product of The Cross-Calibrated Multi-Platform (CCMP). The reanalysis dataset is the GLORYS12V1 reanalysis dataset from the Copernicus Marine Environment Monitoring Service (CMEMS), with a spatial resolution of 1 / 12° and a temporal resolution of days. The vertical spatial distribution of the reanalysis dataset consists of 23 layers, including 0.49 m, 2.65 m, 5.08 m, 7.93 m, 11.41 m, 15.81 m, 21.60 m, 29.44 m, 40.34 m, 55.76 m, 77.85 m, 92.32 m, 109.73 m, 130.67 m, 155.85 m, 186.13 m, 222.48 m, 266.04 m, 318.13 m, 380.21 m, 453.94 m, 541.09 m, and 643.57 m.

[0020] The preprocessing method for the acquired multi-source remote sensing data is as follows: 1. Bilinear data interpolation: The remote sensing images of different resolutions in the multi-source remote sensing data are uniformly processed to a resolution of 1 / 12° and a time resolution of 1 day; 2. Outlier handling: Replace missing data values ​​in multi-source remote sensing data with 0; 3. Normalize multi-source remote sensing data. Standard Scaler can be used to normalize the dataset. The specific steps are as follows: ; in, For normalized multi-source remote sensing data, This represents multi-source remote sensing data after normalization. Represents the mean of all sample data. This represents the standard deviation of all sample data.

[0021] 4. Integrate the remote sensing data and the reanalysis dataset into spatiotemporally correlated tensors, with tensor dimensions of [missing information]. ,in, For batch quantity, For time step, The number of channels in a remote sensing dataset represents the number of variables in the remote sensing image, while the number of channels in a 3D temperature background field dataset represents the number of layers. and It represents the length and width of the forecast sea area.

[0022] A spatiotemporal sequence forecasting model was constructed, and the model was trained using historical ocean remote sensing data and reanalysis data, and then trained until it was qualified.

[0023] For details, see Figure 2 The aforementioned spatiotemporal sequence prediction model includes a first-dimensional reshaping module 1, a convolution module 2, a channel attention module 3, a second-dimensional reshaping module 4, a global position encoding module 5, a multi-scale residual window spatiotemporally coupled attention module 6, a third-dimensional reshaping module 7, a deconvolution module 8, and a fourth-dimensional reshaping module 9.

[0024] The first-dimensional reshaping module 1 is used to reshape tensors. Shape reshaping This is to facilitate the subsequent extraction of spatial features within a single frame image.

[0025] Convolution module 2 is used to reshape the tensor of the first dimension reshaping module 1 through convolution. The encoding is done into a low-dimensional latent space to initially extract spatial feature information, as follows: ; in, This represents the initially extracted spatial feature information. This represents the activation function. This indicates a normalization operation. This represents the convolution operation. This represents the tensor output by the first-dimensional reshaping module.

[0026] Channel attention module 3 is used to emphasize or suppress information between channels, and to generate a vector containing key remote sensing features based on the spatial feature information initially extracted by convolution module 2, as follows: ; in, This generates a vector containing key remote sensing feature information. Indicates channel attention. Indicates spatial attention, and Represents adaptive average pooling and max pooling. This represents the sigmoid function. The translational and overlapping convolutional processes often learn multiple redundant features, and adjacent spatial regions of the feature map contain highly overlapping information. Since the input multi-source remote sensing satellite data have varying impacts on the inversion results, a channel attention mechanism, Convolutional Block Attention Module (CBAM), and global position encoding are introduced to extract the relationship between channels and space. By emphasizing or suppressing information between channels, the parts beneficial to the reconstruction work are further extracted, helping the model to better focus on key remote sensing image features.

[0027] The second-dimensional reshaping module 4 is used to reshape the vector containing key remote sensing feature information generated by the channel attention module 3. This facilitates the stacking of single-frame images on the timeline and the extraction of spatiotemporal features of tensors.

[0028] Global position encoding module 5 is used to in Embedding operations are performed on the vectors containing key remote sensing features after shape reshaping by the second dimension reshaping module 4, with an encoding dimension of 64. After encoding, the spatiotemporal features are respectively coupled with... The encoded values ​​are added together to generate a vector containing global location information, as shown below: ; in, It is a vector containing global position information. This indicates the Embedding operation.

[0029] The multi-scale residual window spatiotemporal coupling attention module 6 is used to extract spatiotemporal features from the vector containing global position information generated by the global position encoding module 5. Specifically, the multi-scale residual window spatiotemporal coupling attention module 6 includes a window segmentation module 61, a first attention mechanism module 62, a window downsampling module 63, a second attention mechanism module 64, a window upsampling module 65, and a window upsampling module 66.

[0030] The window segmentation module 61 segments vectors containing global location information, thereby dividing a high-resolution, large-size image into multiple smaller images and reducing the global computational overhead required by the attention mechanism. The first attention mechanism module 62 extracts the spatiotemporal features of the window from the vectors containing global location information using the attention mechanism. The window downsampling module 63 resamples the spatiotemporal features of the window to expand the receptive field of the attention mechanism. The second attention mechanism module 64 extracts spatiotemporal features from the resampled data using the attention mechanism. The window upsampling module 65 restores the spatiotemporal features extracted by the second attention mechanism module to the window size, and then performs a residual connection with the spatiotemporal features extracted by the first attention mechanism module to obtain spatiotemporal features from multiple windows. The window merging module 66 merges the spatiotemporal features from multiple windows to obtain the spatiotemporal features extracted by the multi-scale residual window spatiotemporally coupled attention module 6.

[0031] The third-dimensional reshaping module 7 is used to reshape the spatiotemporal features extracted by the multi-scale residual window spatiotemporal coupling attention module 6 back to their original shape. .

[0032] The deconvolution module 8 is used to decode the spatiotemporal features reshaped by the third-dimensional reshaping module 7, and performs residual connection with the spatial feature information initially extracted by the convolution module 2 to map the decoded spatiotemporal features onto the required profile depth, as follows: ; in, The mapping result output by deconvolution module 8. This is a deconvolution operation.

[0033] The fourth-dimensional reshaping module 9 is used to reshape the mapping result of the deconvolution module 8 to... Dimensions are used to obtain forecast results.

[0034] During training, the spatiotemporal series prediction model divides the training, validation, and test datasets in an 8:1:1 ratio. Training and validation determine the optimal model parameters, while testing freezes these parameters. The training, validation, and testing processes are implemented using distributed data parallelism (DDP). The model is built using the PyTorch framework with a learning rate of 0.001, the Adam optimizer, the MSE loss function, and an early stopping group set to 30. The MSE is calculated as follows: ; in, , , They are respectively index and This represents the model-predicted temperature and the actual temperature at the location represented by the target index at a time step of T.

[0035] After training, validation can be performed on multi-source remote sensing image variable selection, time step, model comparison, and operational forecasting, as detailed below: (1) Verification of variable selection for multi-source remote sensing images: The variables include SST, SSS, ADT, UWND and VWND. The differences in different forecast results were compared by sequentially removing each sea surface remote sensing data.

[0036] (2) Time step verification: The introduction of historical data enables the model to learn the changing trend in the time series. Therefore, it is necessary to determine the optimal preceding time step through experiments.

[0037] (3) Model comparison and verification: In order to measure the ability of this method to predict ocean temperature structure, the prediction results of this method are compared with those of other spatiotemporal prediction models.

[0038] (4) Operational forecast verification: Under the IV-TT Class 4 standard framework, this method is compared with internationally advanced marine operational forecasting systems.

[0039] The preprocessed multi-source remote sensing data is input into a trained spatiotemporal sequence prediction model to obtain prediction results, which are then visualized.

[0040] In the later stages, error analysis can be performed on the forecast results of the spatiotemporal series forecasting model. The evaluation indicators for error analysis include root mean square error. and coefficient of determination The specific calculation method is as follows: ; ; ; in, denoted as mean squared error, and m is the sample size of the forecast results. The actual value corresponding to the forecast result. The forecast results are from the spatiotemporal series forecasting model. This is the average of the true values.

[0041] See Figure 3This paper demonstrates the root mean square error (RMSE) of end-to-end predictions of the 7-day ocean 3D temperature environment using remote sensing images, comparing the predicted and actual values. The invention compares the predicted and actual values ​​with several other publicly available spatiotemporal series prediction models. The horizontal axis represents the RMSE between the predicted and actual values, and the vertical axis represents depth. As shown in the figure, due to the presence of the thermocline, all models exhibit an increasing error trend in the subsurface layer. With increasing depth, the ocean temperature structure gradually stabilizes, and the prediction error decreases accordingly. Compared to other spatiotemporal series prediction models, the model proposed in this invention achieves the best prediction results.

[0042] The above description is merely a preferred embodiment of the present invention. It should be noted that for those skilled in the art, other parts not specifically described are existing technology or common knowledge. Several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature forecasting method, characterized in that, include: Acquire multi-source remote sensing data and ocean reanalysis data, and preprocess the acquired multi-source remote sensing data and reanalysis data; A spatiotemporal sequence forecasting model is constructed, and the model is trained using historical ocean remote sensing data and reanalysis data, and the model is trained to a qualified state. The preprocessed multi-source remote sensing data is input into the trained spatiotemporal sequence prediction model to obtain prediction results, and the prediction results are then visualized.

2. The artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature forecasting method according to claim 1, characterized in that, The remote sensing data includes sea surface temperature, absolute dynamic height of sea surface, sea surface salinity, zonal wind speed, and meridional wind speed; the reanalysis dataset is a daily temperature dataset with a resolution of 1 / 12°, and its vertical spatial distribution is multi-layered.

3. The artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature forecasting method according to claim 2, characterized in that, The reanalysis dataset has a vertical spatial distribution of 23 layers, including 0.49m, 2.65m, 5.08m, 7.93m, 11.41m, 15.81m, 21.60m, 29.44m, 40.34m, 55.76m, 77.85m, 92.32m, 109.73m, 130.67m, 155.85m, 186.13m, 222.48m, 266.04m, 318.13m, 380.21m, 453.94m, 541.09m and 643.57m.

4. The artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature forecasting method according to claim 2, characterized in that, The preprocessing method is as follows: Remote sensing images of different resolutions from multi-source remote sensing data are uniformly processed to a 1 / 12° resolution and a time resolution of 1 day. Replace missing data values ​​in multi-source remote sensing data with 0; Normalize multi-source remote sensing data; The remote sensing data and the reanalysis dataset were respectively integrated into spatiotemporally correlated tensors, with tensor dimensions of [missing information]. ,in, For batch quantity, For time step, For the number of channels, and It represents the length and width of the forecast sea area.

5. The artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature forecasting method according to claim 4, characterized in that, The spatiotemporal sequence prediction model includes: The first-dimensional reshaping module is used to reshape tensors. Shape reshaping ; A convolution module is a tensor used to reshape the module by convolution. The encoding is done into a low-dimensional latent space to initially extract spatial feature information, as follows: ； in, This represents the initially extracted spatial feature information. This represents the activation function. This indicates a normalization operation. This represents the convolution operation. This represents the tensor output by the first-dimensional reshaping module; The channel attention module is used to emphasize or suppress information between channels and generate vectors containing key remote sensing features based on the spatial feature information initially extracted by the convolution module, as detailed below: ； in, This generates a vector containing key remote sensing feature information. Indicates channel attention. Indicates spatial attention, and Represents adaptive average pooling and max pooling. Represents the sigmoid function; The second-dimensional reshaping module is used to reshape the vector containing key remote sensing feature information generated by the channel attention module. ; Global position encoding module, used to in In terms of dimensions, the vectors containing key remote sensing features after shape reshaping by the second-dimensional reshaping module are encoded with 64 dimensions. After encoding, the spatiotemporal features are then compared with... The encoded values ​​are added together to generate a vector containing global location information, as shown below: ； in, It is a vector containing global position information. Indicates encoding operation; A multi-scale residual window spatiotemporally coupled attention module is used to extract spatiotemporal features from vectors containing global position information generated by the global position encoding module; The third-dimensional reshaping module is used to reshape the spatiotemporal features extracted by the multi-scale residual window spatiotemporally coupled attention module back to their original shape. ; The deconvolution module is used to decode the spatiotemporal features reshaped by the third-dimensional reshaping module and perform residual connection with the spatial feature information initially extracted by the convolution module to map the decoded spatiotemporal features onto the required profile depth, as follows: ； in, The mapping result output by the deconvolution module. This is a deconvolution operation; The fourth-dimensional reshaping module is used to reshape the mapping result of the deconvolution module to... Dimensions are used to obtain the forecast results.

6. The artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature forecasting method according to claim 5, characterized in that, The multi-scale residual window spatiotemporally coupled attention module includes: The window segmentation module is used to segment vectors containing global position information; The first attention mechanism module is used to extract the spatiotemporal features of the window from a vector containing global position information using the attention mechanism; The window downsampling module is used to resample the spatiotemporal features of the window in order to expand the receptive field of the attention mechanism. The second attention mechanism module is used to extract spatiotemporal features from the resampled data using the attention mechanism; The window upsampling module is used to restore the spatiotemporal features extracted by the second attention mechanism module to the window size, and then perform residual connection with the spatiotemporal features extracted by the first attention mechanism module to obtain spatiotemporal features of multiple windows. The window merging module is used to merge the spatiotemporal features of multiple windows to obtain the spatiotemporal features extracted by the multi-scale residual window spatiotemporally coupled attention module.

7. The artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature forecasting method according to claim 1, characterized in that, The spatiotemporal sequence prediction model was trained using the PyTorch framework with a learning rate of 0.001, an optimizer of Adam, a loss function of MSE, and an early stopping group of 30.

8. The artificial intelligence-based remote sensing end-to-end three-dimensional ocean temperature forecasting method according to claim 1, characterized in that, Also includes: Error analysis is performed on the forecast results of the spatiotemporal series forecasting model. The evaluation index of the error analysis includes the root mean square error. and coefficient of determination The specific calculation method is as follows: ； ； ； in, denoted as mean squared error, and m is the sample size of the forecast results. The actual value corresponding to the forecast result. The forecast results are from the spatiotemporal series forecasting model. This is the average of the true values.

Images