A method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM

Abstract

The present invention discloses a method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM, and relates to the field of ocean remote sensing technology. The present invention proposes a method for predicting ocean surface temperature that integrates an OI method and a ConvLSTM model. The OI method solves the spatial integrity problem of data and provides high-quality input for the ConvLSTM. The ConvLSTM model, based on the data processed by the OI method, leverages its ability to learn spatiotemporal features to achieve accurate time series prediction. The two methods work together to ensure the spatial continuity and physical rationality of sea surface temperature data while effectively predicting its changes in time series. This provides a more reliable and accurate solution for sea surface temperature prediction in complex ocean environments where data acquisition is difficult.

Description

Technical Field

[0001] The present invention relates to the field of ocean remote sensing detection technology, and in particular to a method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM. Background Art

[0002] Sea surface temperature (SST) is the temperature at the interface between the ocean and the atmosphere. As a core variable in air-sea interactions, its changes directly influence atmospheric circulation, extreme weather events (such as typhoons and heat waves), and the stability of marine ecosystems. Accurate SST prediction is crucial for climate research, disaster warning, and fishery resource management.

[0003] In recent years, with advances in satellite remote sensing technology, remote sensing observations have become the primary means of obtaining SST. Currently, satellite SST observations rely primarily on infrared and microwave sensors. Infrared sensors offer the advantage of high spatial resolution, but are susceptible to interference from aerosols and atmospheric clouds. Microwave sensors, on the other hand, possess strong cloud penetration capabilities, enabling all-day, all-weather observations, but their spatial resolution is relatively low. Given the limitations of single-sensor observation capabilities, satellite multi-source data prediction technology has emerged. This technology aims to integrate the strengths of different sensors to enhance the temporal and spatial integrity and accuracy of SST products, better meeting practical application needs.

[0004] Optimal interpolation (OI) is a method for predicting SST data. This method fully considers the spatial distribution characteristics of each pixel, interpolating unevenly distributed data onto corresponding grid points to obtain the optimal estimate. This effectively addresses the problem of sparse spatial distribution of ocean data. However, traditional OI methods fail to fully consider the spatiotemporal correlation of observed data during background field construction and weighting, limiting their applicability and accuracy in highly dynamic ocean regions.

[0005] In recent years, the application of deep learning in the field of marine remote sensing has gradually emerged. Among them, the Convolutional Long Short-Term Memory (ConvLSTM) network is suitable for processing remote sensing data with strong spatiotemporal dependencies because it has both the spatial feature extraction capabilities of convolutional neural networks and the temporal modeling capabilities of LSTM.

[0006] Given this, the challenge facing existing technologies is how to combine optimal interpolation methods with ConvLSTM to fully leverage the advantages of OI in constructing a physically consistent background field and suppressing observation noise, while also leveraging the ConvLSTM's spatiotemporal learning capabilities to dynamically capture the spatiotemporal evolution of SST, thereby improving the accuracy and stability of predictions in areas without measurements. This combination of methods is expected to overcome the limitations of traditional OI methods in highly dynamic ocean regions and enhance the accuracy and stability of SST data predictions.

[0007] The information disclosed in this background section is only intended to enhance understanding of the overall background of the invention and should not be considered as an admission or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the Invention

[0008] In response to the above technical problems, an embodiment of the present invention provides an ocean surface temperature prediction method based on optimal interpolation combined with ConvLSTM to solve the problems raised in the above background technology.

[0009] The present invention provides the following technical solution: a method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM, comprising the following steps:

[0010] Step 1: Acquire AMSR2, MODIS, HY1C, HY1D satellite data, OSTIA reanalysis data, and iQuam buoy data in sequence; construct a multi-source sea surface temperature dataset;

[0011] Step 2: Perform quality control (preprocessing) on ​​the multi-source sea surface temperature dataset to remove abnormal or invalid sea surface temperature data;

[0012] Step 3: Using the OSTIA reanalysis data as the initial background field, the optimal interpolation algorithm is used to interpolate the sea surface temperature data of AMSR2, MODIS, HY1C, and HY1D in sequence to generate a standard gridded sea surface temperature product;

[0013] Step 4: Divide the optimal interpolated sea surface temperature grid product into a training set, a validation set, and a test set, and perform normalization.

[0014] Step 5: Build an OI-ConvLSTM model and use the normalized training set and validation set to train the model and obtain the optimal OI-ConvLSTM model.

[0015] Step 6: Use the optimal OI-ConvLSTM model to predict the sea surface temperature of the normalized test set and perform denormalization to obtain the actual sea surface temperature prediction data;

[0016] Step 7: Use iQuam buoy data to evaluate the accuracy of the predicted sea surface temperature product.

[0017] Preferably, the quality control method includes: reading the sea surface temperature variables in AMSR2, MODIS, HY1C, and HY1D, eliminating sea surface temperature values ​​less than 0°C or greater than 40°C, setting the -999.0 and -32767.0 fill values ​​to 0, and using a variable representing the quality level to mask pixel values ​​with poor quality.

[0018] Preferably, in step 3, the optimal interpolation algorithm mainly includes: background field error covariance matrix B and observation field error covariance matrix R.

[0019] Preferably, the background field error covariance matrix, the diagonal matrix D composed of the background field errors and the background field horizontal correlation matrix Composition; the formula is: Among them, D is the diagonal matrix composed of background field errors, is the horizontal correlation matrix of the forecast background field, a and b are the correlation scales in longitude and latitude respectively, are the relevant distances in longitude and latitude directions respectively.

[0020] Preferably, the observation field error covariance matrix is ​​calculated using the same method as the background field error covariance matrix; the calculation formula is: Among them, i and j represent different observation points.

[0021] Preferably, in step 4, the sea surface temperature gridded product after optimal interpolation is divided into a training set, a validation set, and a test set in a ratio of 7:1:2;

[0022] The model's input data dimensions are 5, including: batch size, time step, number of channels, latitude, and longitude;

[0023] The label data output dimensions are 4, including: batch size, latitude value, longitude value, and channel size.

[0024] Preferably, in step 4, the normalization formula is as follows: ;

[0025] Where, is the normalized value, is the maximum sea surface temperature in the sample, is the minimum sea surface temperature in the sample, is the true sea surface temperature value in the sample.

[0026] Preferably, the OI-ConvLSTM model structure includes: three layers of stacked ConvLSTM; each layer uses a 3×3 convolution kernel and 64 filters to capture the spatiotemporal evolution characteristics of SST through temporal information; each layer of the ConvLSTM model includes: an input gate, a forget gate, a memory unit, and an output gate.

[0027] Specifically, the input gate, forget gate, memory unit, and output gate are expressed as: in, For the input gate, For the Forgotten Gate, For memory units, For the output gate, is the sigmoid activation function, is the hyperbolic tangent activation function, is the input of the current time step, and are the hidden states at time t and t-1 respectively, is the memory unit of the previous time step, W and b are the weight matrix and bias matrix parameters respectively, represents product and * represents convolution operation.

[0028] Preferably, in step 6, the predicted sea surface temperature is denormalized to obtain the actual sea surface temperature prediction data; the denormalization formula is: ;

[0029] Where, is the true sea surface temperature value in the sample, is the normalized value, is the maximum sea surface temperature in the sample, is the minimum sea surface temperature in the sample.

[0030] Preferably, in step 7, the accuracy of the predicted sea surface temperature product is evaluated, and the accuracy evaluation indicators include: correlation coefficient , mean relative error , mean deviation , mean standard deviation The specific formula is: Among them, X is the observed value, X ’ is the reference value, and N is the number of matching data points.

[0031] The embodiment of the present invention provides a method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM, which has the following beneficial effects:

[0032] (1) This paper proposes a sea surface temperature prediction method that integrates the OI method and the ConvLSTM model. The OI method solves the spatial integrity problem of data and provides high-quality input for ConvLSTM. The ConvLSTM model, based on the data processed by the OI method, exerts its ability to learn spatiotemporal features and achieves accurate time series prediction.

[0033] (2) The two work together to ensure that this method can not only ensure the spatial continuity and physical rationality of sea surface temperature data, but also effectively predict its changes in time series. In the case of complex ocean environments and difficult data acquisition, it provides a more reliable and accurate solution for sea surface temperature prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flowchart of a method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM, provided by an embodiment of the present invention;

[0035] Figure 2 1 is a structural diagram of an optimal interpolation algorithm model according to an embodiment of the present invention;

[0036] Figure 3 This is a structural diagram of the OI-ConvLSTM algorithm model described in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0038] The design principle of the ocean surface temperature prediction method based on optimal interpolation combined with ConvLSTM in the present invention is as follows:

[0039] (1) Limitations of multi-source data: When satellite remote sensing is used to obtain sea surface temperature (SST), infrared sensors have high resolution but are easily affected by cloud cover. Microwave sensors can observe data in all weather conditions but have low resolution, resulting in spatial discontinuity and uneven distribution of data. For example, in areas covered by clouds, infrared sensors cannot obtain valid data, resulting in data loss.

[0040] (2) The role of optimal interpolation (OI): The OI method uses the OSTIA reanalysis data as the initial background field. The OSTIA data itself has integrated multiple data and has a certain spatial continuity and physical basis. On this basis, the OI method calculates the background field error covariance matrix and the observation field error covariance matrix, fully considers the spatial distribution characteristics of each pixel, and interpolates the unevenly distributed data in multi-source observation data such as AMSR2 and MODIS to the corresponding grid points according to certain weights; thereby constructing a physically consistent and spatially complete background field, effectively alleviating the spatial missing problem in multi-source remote sensing data, and making the originally discrete and discontinuous data form a continuous gridded SST product.

[0041] (3) Characteristics of ConvLSTM: ConvLSTM combines the spatial feature extraction capability of convolutional neural networks (CNN) with the temporal modeling capability of long short-term memory networks (LSTM). When processing sea surface temperature data, its convolutional structure can extract the spatial characteristics of SST data, such as the shape of the temperature distribution, gradient changes, etc. The LSTM structure is good at processing time series data, can remember the information of past time steps, and update the memory according to the current input, thereby capturing the evolution of SST over time, such as seasonal changes, short-term fluctuations, etc.

[0042] (4) Advantages of ConvLSTM combined with OI data: The spatially continuous gridded SST products generated by the OI method are used as the input of ConvLSTM, providing the model with high-quality, continuous spatiotemporal data; ConvLSTM learns and trains based on these data, which can better explore the dependency of SST in the time dimension and achieve accurate prediction of SST at future time points, thereby giving the method better time series prediction capabilities.

[0043] 2. In response to the problems mentioned in the above background technology, the embodiments of the present invention provide an ocean surface temperature prediction method based on optimal interpolation combined with ConvLSTM to solve the above technical problems. The technical solution is as follows:

[0044] The following is combined with Figure 1-3 , and specific implementation methods are further described to illustrate the present invention.

[0045] First, install Python programming software on your terminal. This requires Python 3.10, tensoR, flow 2.16.1, and keras 3.0.5. TensoR, flow, and keras are open-source software libraries for deep learning. TensoR and flow provide a flexible platform for building, training, and deploying complex neural network models in multiple programming languages. Keras is an advanced neural network API that provides a simpler and more user-friendly interface based on tensoR and flow, allowing users to quickly build and run common neural network models.

[0046] Figure 1 As shown in the figure, a method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM is provided; the method comprises the following steps:

[0047] Step S1, obtain AMSR2, MODIS, HY1C, HY1D satellite data, OSTIA reanalysis data, iQuam buoy data, and construct a multi-source sea surface temperature dataset;

[0048] Step S2: performing quality control on the acquired multi-source sea surface temperature dataset;

[0049] Step S3: using the OSTIA reanalysis data as the initial background field, the preprocessed multi-source sea surface temperature dataset is interpolated using the optimal interpolation method to generate a standard gridded sea surface temperature product.

[0050] Step S4: dividing the sea surface temperature grid product after optimal interpolation into a training set, a validation set, and a test set;

[0051] Step S5: Construct a ConvLSTM model and use the training set and validation set to train the model to obtain the optimal ConvLSTM model;

[0052] Step S6: Use the optimal ConvLSTM model to predict the sea surface temperature using the test set, and use the iQuam buoy data to evaluate the accuracy of the predicted sea surface temperature product.

[0053] In this embodiment, in step S1, AMSR2, MODIS, HY1C, HY1D satellite data, OSTIA reanalysis data, and iQuam buoy data are obtained to construct a multi-source sea surface temperature dataset. The details are as follows:

[0054] Regarding MODIS satellite data, MODIS is a Moderate Resolution Imaging Spectroradiometer (MODIS) instrument carried on the AQUA and TERRA satellites. TERRA is a morning satellite with a transit time of approximately 10:30 AM, while AQUA is an afternoon satellite with a transit time of approximately 1:30 PM. It captures data in 36 spectral bands at various spatial resolutions. SST products are generated using an improved nonlinear SST algorithm in the 4, 11, and 12 μm infrared bands. The MODIS sea surface temperature data used in this paper is sourced from Ocean Color (https: / / oceancolor.gsfc.nasa.gov). The product level is L3 mapped data with a spatial resolution of 4 km. This product is calibrated and radiometrically corrected global single-day data, projected to the WGS84 coordinate system, and subjected to rigorous quality control.

[0055] Regarding AMSR-2 satellite data, AMSR-2 is an advanced microwave scanning radiometer aboard the JAXA Global Change Observation Mission-Water (GCOM-W) satellite. Its transit time is 1:30 p.m., and its antenna rotates every 1.5 seconds, capturing data across a 1,450-kilometer swath. Every two days, it provides daytime and nighttime data covering over 99% of the Earth. The AMSR-2 sea surface temperature data used in this paper comes from the US Remote Sensing Systems (RSS), with a product level of L3 and a spatial resolution of 0.25° x 0.25°.

[0056] Regarding the HY1C and HY1D satellite data, HY1C / 1D are operational ocean remote sensing satellites from my country's HY-1 series. They carry the COCTS ocean color scanner, which features two thermal infrared channels for sea surface temperature (SST) observations. The HY1C / 1D SST data used in this paper is from the National Satellite Ocean Application Center (https: / / osdds.nsoas.org.cn / ), available free of charge. The product level is L3A, with a spatial resolution of 4 km.

[0057] iQuam data, developed by NOAA, includes data from Argo floats, conventional drifting, high-resolution drifting, tropical moored, coastal moored, coral reef observation buoys, conventional ships, and International Maritime Organization (IMO) vessels. iQuam data is stored in NetCDF4 format, as a monthly data file, updated every 24 hours. The data file is divided into two parts: global attributes and data variables. This paper uses data from observation types of ship, drifting buoy, tropical moored buoy, coastal moored buoy, and high-resolution drifting buoy to verify the accuracy of the fused product.

[0058] OSTIA (Sea Surface Temperature and Ice Analysis) data is a global sea surface temperature and sea ice analysis product developed by the UK Met Office. It combines satellite data (AVHRR, AMSR, TMI, AATSR, and SEVIRI) with in situ observations to produce high-resolution daily sea surface temperature and sea ice area fraction analysis data with a grid spacing of 1 / 20°. The data is available from https: / / www.ncei.noaa.gov / data / oceans / ghrsst / L4 / GLOB / UKMO / OSTIA / .

[0059] In this embodiment, in step S2, quality control is performed on the acquired multi-source sea surface temperature dataset. The specific method of quality control is as follows:

[0060] This dataset is based on L3 sea surface temperature data from AMSR2, MODIS, HY1C, and HY1D, covering the period from June 1, 2024, to December 31, 2024, and the spatial range from 104°E to 122°E, and 10°N to 26°N. First, we read the sea surface temperature variable from each data file and remove sea surface temperature values ​​less than 0°C or greater than 35°C. Then, we set the fill values ​​of -999.0 and -32767.0 to 0. Next, we mask out poor-quality pixels using a variable representing the quality level, ultimately generating a high-quality, single-day, multi-source sea surface temperature dataset.

[0061] In this embodiment, in step S3, the preprocessed multi-source sea surface temperature dataset is interpolated using the OSTIA reanalysis data as the initial background field, using the optimal interpolation method to generate a standard gridded sea surface temperature product. The optimal interpolation method is an objective analysis method. During the analysis process, the analysis value at the spatial grid point is obtained by weighting the deviation of the observation point relative to the background field. The weight coefficient should minimize the analysis error of the grid point and is not arbitrarily selected. The specific method is as follows:

[0062] This paper uses preprocessed 0.1° x 0.1° OSTIA reanalysis data as the initial background field and 0.1° x 0.1° AMSR2 sea surface temperature data as the observation field. The study area is 104°E-122°E and 10°N-26°N. Optimal interpolation and fusion are performed to finally generate a 0.1° x 0.1° sea surface temperature product (OI-SST) for sea surface temperature prediction. The specific formula for optimal interpolation is: ;

[0063] in, Represents the analytical value of the variable at the spatial grid point, represents the background field value of the variable at the spatial grid point, is the weight coefficient, represents the observed value of the variable at the observation point, represents the background field value of the variable at the observation point, is the observation operator. Weight coefficient matrix The expression is: ;

[0064] in, represents the transpose of a matrix, represents the background field error covariance matrix, Represents the observation error covariance matrix. Calculate the weight coefficient , we need to estimate the background field error covariance matrix first and the observation error covariance matrix .

[0065] (1) Calculate the background field error covariance matrix

[0066] Background field error covariance matrix, diagonal matrix D composed of background field errors and background field horizontal correlation matrix The composition formula is: The composition formula is: Among them, D is the diagonal matrix composed of background field errors, is the horizontal correlation matrix of the forecast background field, a and b are the correlation scales in longitude and latitude respectively, are the relevant distances in longitude and latitude directions respectively.

[0067] (2) Calculate the observation field error covariance matrix The same method as the background field error covariance matrix is ​​used for the observation field error covariance matrix. It is generally believed that the observation errors at different positions are not related, that is, the calculation formula is: Among them, i and j represent different observation points.

[0068] (3) Calculate weight According to the background field error covariance matrix and the observation field error covariance matrix The optimal interpolation coefficient matrix is ​​obtained by solving the equation , and calculate the sea surface temperature after the fusion of OSTIA and AMSR2.

[0069] Subsequently, the result of this optimal interpolation is used as the background field for the next optimal interpolation. The sea surface temperature data of MODIS, HY1C, and HY1D are introduced as the observation field in turn. The optimal interpolation process is repeated to gradually optimize the spatial distribution and accuracy of the sea surface temperature field and obtain the final sea surface temperature fusion product.

[0070] In this embodiment, in step S4, the sea surface temperature grid product after optimal interpolation is divided into a training set, a validation set, and a test set, and normalized. The details are as follows:

[0071] First, we selected OI-SST data from January 1, 2024, to December 31, 2024, and divided it into training, validation, and test sets in a ratio of 7:1:2. The model's input data has five dimensions: batch size B, time step T, number of channels C, latitude H, and longitude W. The label data output has four dimensions: batch size B, channel size C, latitude H, and longitude W. This ensures that the prediction effect of five consecutive time series is output as one time series.

[0072] Secondly, the divided data set is normalized before being input into the model. The purpose is to accelerate convergence, avoid gradient vanishing, prevent feature dominance effect, improve numerical stability, and better adapt to the network weight initialization. The formula is as follows: Where, is the normalized value, is the maximum sea surface temperature in the sample, is the minimum sea surface temperature in the sample, is the true sea surface temperature value in the sample.

[0073] Specifically, in step S5, an OI-ConvLSTM model is constructed, and the model is trained using the training set and the validation set to obtain the optimal OI-ConvLSTM model. The details are as follows: (1) Constructing an OI-ConvLSTM model

[0074] like Figure 3 As shown in the figure, the OI-ConvLSTM model developed in the Python 3.10 environment uses the PyTorch framework. The input is the optimally interpolated sea surface temperature (OI-SST). The model consists of three stacked ConvLSTM layers, each using a 3×3 convolution kernel and 64 filters to capture the spatiotemporal evolution of SST using temporal information. Each ConvLSTM layer consists of an input gate, a forget gate, a memory unit, and an output gate. The expression is: in, For the input gate, For the Forgotten Gate, For memory units, For the output gate, is the sigmoid activation function, is the hyperbolic tangent activation function, is the input of the current time step, and are the hidden states at time t and t-1 respectively, is the memory unit of the previous time step, W and b are the weight matrix and bias matrix parameters respectively, represents product and * represents convolution operation.

[0075] (2) Model training

[0076] The divided training set is used as the input of the OI-ConvLSTM model, with a tensor size of (B, T, C, H, W). The network output is a single SST image predicted for the next moment (day T+1), with a tensor size of (B, C, H, W). The model uses the Adaptive Moment Estimation (Adam) optimization algorithm during training, and chooses the mean squared error (MSE) as the loss function. The training loss is recorded for each epoch. Training is terminated when the loss function does not improve, and the resulting model is used as the optimal model for sea surface temperature prediction. The formula for the mean squared error is: in, is the true value of the i-th data in a batch, is the predicted value given by the model.

[0077] Other parameter settings are as follows: time series length seq_length = 7, prediction length pred_len = 1, batch size batch_size = 4, number of training rounds num_epochs = 500, learning rate learning_rate = 0.001.

[0078] Specifically, in step S6, the optimal ConvLSTM model is used to predict the sea surface temperature using the test set, and the accuracy of the predicted sea surface temperature product is evaluated using the iQuam buoy data. The specific steps of the accuracy evaluation are as follows:

[0079] (1) Model prediction

[0080] First, the divided test set is used as the input of the optimal ConvLSTM model to obtain the normalized predicted sea surface temperature. Then, the denormalization operation is performed and the projection output is performed to obtain the final sea surface temperature data. The denormalization formula is: Where, is the true sea surface temperature value in the sample, is the normalized value, is the maximum sea surface temperature in the sample, is the minimum sea surface temperature in the sample.

[0081] (2) Accuracy evaluation

[0082] The accuracy of the predicted sea surface temperature was tested using iQuam data. The spatiotemporal matching principles are as follows: ① The time window is set to the current day; ② The spatial window is 3×3 pixels, and the average of the 3×3 pixels adjacent to the sampling point is taken as the sea surface temperature value at that point; ③ The number of valid pixels in the window is greater than 50%; ④ Pixels with values ​​outside of X±(3×σ) are eliminated; ⑤ The coefficient of variation (CV) is calculated as CV=σ / X. If CV≤0.15, the uniformity is determined. The following four evaluation indicators are used: correlation coefficient r, mean relative error (MARE), mean bias (Bias), and mean standard deviation (STD). Their calculation formulas are: Among them, Bias is the mean bias, STD is the mean standard deviation, r is the correlation coefficient, MARE is the mean relative error, X is the observed value, X ’ is the reference value, and N is the number of matching data points.

[0083] 2. Test Results

[0084] Based on the optimal sea surface temperature fusion method obtained through the above training, this test selected AMSR2, MODIS, HY1C, and HY1D satellite data from January 2024 within the range of 104°E-122°E and 10°N-26°N, as well as OSTIA reanalysis data for fusion. The accuracy of the sea surface temperature fusion results was evaluated using iQuam buoy data from January 2024. Following the above-mentioned spatiotemporal matching rules, the final matching data points between the iQuam buoy data and the sea surface temperature fusion data were 188,183, with a correlation coefficient r of 0.99, an average relative error of 1.45%, an average deviation of -0.21°C, and a standard deviation of 0.48°C. This shows that the sea surface temperature fusion product obtained using the sea surface temperature fusion algorithm provided by the present invention has high accuracy and meets business needs.

[0085] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM, characterized in that: The following steps are involved: Step 1: Acquire AMSR2, MODIS, HY1C, HY1D satellite data, OSTIA reanalysis data, and iQuam buoy data in sequence; construct a multi-source sea surface temperature dataset; Step 2: Perform quality control on the multi-source sea surface temperature dataset and remove abnormal or invalid sea surface temperature data; Step 3: Using the OSTIA reanalysis data as the initial background field, the optimal interpolation algorithm is used to interpolate the sea surface temperature data of AMSR2, MODIS, HY1C, and HY1D in sequence to generate a standard gridded sea surface temperature product; Step 4: Divide the optimal interpolated sea surface temperature grid product into a training set, a validation set, and a test set, and perform normalization. Step 5: Build an OI-ConvLSTM model and use the normalized training set and validation set to train the model and obtain the optimal OI-ConvLSTM model. The OI-ConvLSTM model structure consists of three stacked ConvLSTM layers. Each layer uses a 3×3 convolution kernel and 64 filters to capture the spatiotemporal evolution characteristics of SST through temporal information. Each ConvLSTM layer includes an input gate, a forget gate, a memory unit, and an output gate. Step 6: Use the optimal OI-ConvLSTM model to predict the sea surface temperature of the normalized test set and perform denormalization to obtain the actual sea surface temperature prediction data; Step 7: Use iQuam buoy data to evaluate the accuracy of the predicted sea surface temperature product; In step 3, the optimal interpolation algorithm mainly includes: background field error covariance matrix B, observation field error covariance matrix R; The background field error covariance matrix B is composed of the diagonal matrix D composed of the background field errors and the background field horizontal correlation matrix ρ; the formula expression is: B=D 0.5 ρD 0.5 ; Where D is the diagonal matrix composed of background field errors, ρ is the horizontal correlation matrix of the forecast background field, a and b are the correlation scales in longitude and latitude respectively, r a 、r b are the relevant distances in longitude and latitude directions respectively; r and t represent placeholders; The observation field error covariance matrix R uses the same method as the background field error covariance matrix B; the calculation formula is: Among them, i, j represent different observation points; According to the observation field error covariance matrix R and the background field error covariance matrix B, the weight coefficient K is solved as follows: K=BH T (HBH T +R) -1 ; Where T represents the transpose of the matrix and H is the observation operator; The specific formula for optimal interpolation is: in, Represents the analytical value of the variable at the spatial grid point, represents the background field value of the variable at the spatial grid point, represents the observed value of the variable at the observation point, Represents the background field value of the variable at the observation point; The result of this optimal interpolation is used as the background field for the next optimal interpolation. The sea surface temperature data of MODIS, HY1C, and HY1D are introduced as the observation field in turn. The optimal interpolation process is repeated to gradually optimize the spatial distribution and accuracy of the sea surface temperature field and obtain the final sea surface temperature fusion product.

2. The method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM according to claim 1, characterized in that: In step 2, the quality control method includes: reading the sea surface temperature variables from AMSR2, MODIS, HY1C, and HY1D, removing sea surface temperature values ​​less than 0°C or greater than 40°C, setting the fill values ​​of -999.0 and -32767.0 to 0, and using a variable representing the quality level to mask pixel values ​​with poor quality.

3. The method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM according to claim 1, characterized in that: In step 4, the optimal interpolated sea surface temperature gridded product is divided into training set, validation set, and test set in a ratio of 7:1:2; The model's input data dimensions are 5, including: batch size, time step, number of channels, latitude, and longitude; The label data output dimensions are 4, including: batch size, latitude value, longitude value, and channel size.

4. The method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM according to claim 1, characterized in that: In step 4, the normalization formula is as follows: Where, SST X is the normalized value, SST max is the maximum sea surface temperature in the sample, SST min is the minimum sea surface temperature in the sample, SST x is the true sea surface temperature value in the sample.

5. The method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM according to claim 1, characterized in that: In step 6, the predicted sea surface temperature is denormalized to obtain the actual sea surface temperature prediction data; the denormalization formula is: SST x =SST X *(SST max -SST min )+SST min ; Where, SST x is the true sea surface temperature value in the sample, SST X is the normalized value, SST max is the maximum sea surface temperature in the sample, SST min is the minimum sea surface temperature in the sample.

6. The method for predicting ocean surface temperature based on optimal interpolation combined with ConvLSTM according to claim 1, characterized in that: In step 7, the accuracy of the predicted sea surface temperature product is evaluated. The accuracy evaluation indicators include: correlation coefficient r, mean relative error MARE, mean bias Bias, and mean standard deviation STD. The specific formula is: Among them, X is the observed value, X ’ is the reference value, and N is the number of matching data points.

Images