Nearshore sea surface temperature fusion method based on deep learning driven variational analysis

Abstract

The present invention discloses a nearshore sea surface temperature fusion method based on deep learning driven variational analysis, which belongs to the field of data processing technology. Specifically, the method comprises: inputting historical sea surface temperature data after variational analysis into a deep learning sea surface temperature prediction model, and outputting the predicted sea surface temperature for a specified future period as the prediction background field of the current variational analysis; establishing a relationship between multi-step sea surface temperature prediction differences and prediction errors based on the deep learning sea surface temperature prediction model, generating a background error variance during fusion, and fusing a spatial distance function with a short-term time correlation function to construct a background error covariance model during variational analysis; assimilating real-time collected sea surface temperature observation data into the predicted background field, combining the background error covariance model and the observation error weight, solving the optimal analysis field and adjusting the resolution, and outputting a nearshore sea surface temperature analysis field; the present invention provides a nearshore sea surface temperature analysis field with both high precision and fine-scale characteristics.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a nearshore sea surface temperature fusion method based on deep learning driven variational analysis. Background Art

[0002] Sea surface temperature (SST) is a core parameter reflecting the dynamic and thermal interactions between the ocean and the atmosphere. Its precise distribution directly influences the global ocean heat budget, the evolution of weather systems, and the functioning of marine ecosystems. High-resolution SST data are essential for numerical forecasting, ecological assessment, and resource management in areas such as climate monitoring, coral reef management, and fisheries research. With the advancement of satellite remote sensing technology, infrared and microwave sensors can now acquire sub-kilometer-level observational data. However, nearshore areas are affected by complex terrain, sparse observations, and environmental noise, and the fusion of multi-source data still faces the challenge of constructing high-precision, high-resolution analysis fields.

[0003] Existing sea surface temperature fusion analysis mainly uses the fusion results of the previous cycle or sea surface temperature climate data as the background field, and adopts the optimal difference or variational method to fuse with the observation field. Nearshore observation data is sparse. When satellite observation data is missing at a certain point at the time of fusion, the analysis information can only be provided by the background field data. Selecting the analysis results of the previous fusion cycle as the background field will bring about the time lag problem of sea surface temperature data. In particular, when a certain area has been missing satellite or measured data for a period of time, the deviation of the background field data will be greater. Existing technologies have significant limitations in nearshore sea surface temperature analysis, and the resolution of most fusion products exceeds 1 km, making it difficult to characterize short-term nearshore thermal variations. When observations are missing, it is easy to cause error accumulation and smooth transition, resulting in loss of fine-scale features, and unable to meet the needs of nearshore research. Summary of the Invention

[0004] The purpose of this invention is to provide a nearshore sea surface temperature fusion method based on deep learning driven variational analysis to solve the following technical problems:

[0005] The nearshore SST fusion is insufficient in estimating the state of nearshore and small- and medium-scale phenomena, and has low spatial resolution and accuracy.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] The nearshore sea surface temperature fusion method based on deep learning driven variational analysis includes the following steps:

[0008] The historical sea surface temperature data after variational analysis is input into the deep learning sea surface temperature prediction model, and the predicted sea surface temperature for a specified period in the future is output as the prediction background field for the current variational analysis;

[0009] Based on the deep learning sea surface temperature prediction model, the relationship between the multi-step sea surface temperature prediction difference and the prediction error is established to generate the background error variance during fusion. The background error covariance model for variational analysis is constructed by fusing the spatial distance function and the short-term temporal correlation function.

[0010] Collect real-time sea surface temperature observation data, including satellite observation data and in-situ buoy observation data, and perform hierarchical quality control and multi-stage bias correction on the sea surface temperature observation data;

[0011] Assimilating the collected sea surface temperature observation data into the prediction background field, combining the background error covariance model and the observation error weight, and solving the optimal analysis field through variational analysis;

[0012] The grid resolution of the optimal analysis field is adjusted to the sub-kilometer level, and the nearshore sea surface temperature analysis field is output.

[0013] As a further solution of the present invention: the sea surface temperature observation data specifically includes:

[0014] The satellite observation data includes infrared sensor data and microwave sensor data. The infrared sensor data covers visible light and thermal infrared bands. The satellite observation data is accompanied by single-point quality marks, sensor deviation and standard deviation information; the on-site buoy observation data comes from drifting buoys and profile buoys, covering surface to shallow water temperatures.

[0015] As a further solution of the present invention: the hierarchical quality control specifically includes:

[0016] The highest confidence data points were selected based on the quality flags attached to the satellite observation data. The solar zenith angle was calculated for the microwave sensor data, and daytime observation data were excluded. The standard deviations of the satellite observation data and the historical sea surface temperature dataset at the same spatial location were compared, and anomalous data points that deviated by more than a set multiple of the standard deviation were eliminated.

[0017] Verify the equipment type and measurement depth identification of on-site buoy data and retain valid observation values ​​of surface water bodies.

[0018] As a further solution of the present invention: the multi-stage deviation correction specifically includes:

[0019] First, the infrared data day and night identification or solar zenith angle calculation is used to filter the night observation data; secondly, the depth deviation estimate originally attached to the sensor is deducted, and the surface or shallow water temperature is converted to the standard depth temperature; finally, the daily average deviation of each sensor from the on-site buoy and the designated infrared benchmark is calculated within the preset spatial grid, and the global deviation field is generated through Gaussian interpolation and dynamically deducted from the original observation value.

[0020] As a further solution of the present invention: the deep learning sea surface temperature prediction model specifically includes:

[0021] The input is continuous gridded 3D sea surface temperature data, and the spatiotemporal features are extracted through a 3D convolutional layer, combined with long short-term memory units to capture the temporal evolution law;

[0022] A multi-step prediction mechanism is used to synchronously output the forecast fields for multiple days in the future, where the forecast results for the first day are used as the forecast background field for variational analysis.

[0023] Historical satellite fusion data is used in the model training phase, and the loss function constrains both spatial gradient and temporal continuity.

[0024] As a further solution of the present invention: the construction process of the background error covariance model is:

[0025] Run the deep learning sea surface temperature prediction model to generate multiple sets of forecast background fields for different forecast periods. Calculate the sum of squares of the differences between each forecast background field and the forecast mean field, and then map it to a spatial error variance field after time series statistics.

[0026] The spatial Gaussian correlation function of fixed scale and the Pearson correlation coefficient of multi-day sliding window are integrated with equal weight coefficients, and the weighted superposition is used to form a spatiotemporal adaptive correlation structure.

[0027] The error variance field and the dynamic correlation structure are combined into a complete background error covariance model according to the covariance decomposition principle.

[0028] As a further solution of the present invention: the process of solving the optimal analysis field is:

[0029] An incremental cost function is constructed, and the difference between the analysis field and the predicted background field is defined as a variable. The background error term is constrained by the background error covariance model, and the observation error term is set as a diagonal matrix with element values ​​taken from the native standard deviation of the sensor. The cost function is minimized through an iterative optimization algorithm to obtain the analysis increment and superimpose it on the predicted background field to generate the optimal analysis field.

[0030] As a further solution of the present invention: when satellite observation data is missing, the optimal analysis field is completely dependent on the predicted background field, and the local spatial constraints are enhanced by the short-time correlation components in the background error covariance model; the physical laws implicit in the deep learning sea surface temperature prediction model are used to maintain the spatiotemporal continuity of the temperature field.

[0031] As a further solution of the present invention: the output nearshore area sea surface temperature analysis adopts grid rule resampling, covering the spatial range from the coastline to the edge of the continental shelf, and the grid resolution is set to sub-kilometer level; the output data includes the sea surface temperature value and the corresponding error estimation field.

[0032] Beneficial effects of the present invention:

[0033] The present invention generates a dynamic variational analysis background field through a deep learning sea surface temperature prediction model, solving the problem that traditional background fields cannot capture the dynamic changes of nearshore thermal structures. Combined with the background error variance calculation based on the predicted spatiotemporal differences and the covariance model that integrates the spatial Gaussian function and short-term correlation, the spatiotemporal adaptive adjustment of the background error covariance is achieved, overcoming the defect of the existing model in insufficient characterization of nearshore short-scale thermal variations; the processing accuracy of nearshore observation data is improved through hierarchical quality control and multi-stage bias correction, and the deep learning sea surface temperature prediction model implicit in the physical laws and the short-time correlation component of the covariance model is used to maintain the spatiotemporal continuity of the temperature field and suppress error accumulation when observations are missing; the grid resolution is optimized to sub-kilometer level, fully releasing the high-resolution potential of satellite infrared data, and providing nearshore areas with a sea surface temperature analysis field with both high precision and fine-scale characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The present invention will be further described below with reference to the accompanying drawings.

[0035] Figure 1 It is a schematic flow diagram of the present invention. DETAILED DESCRIPTION

[0036] 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 ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0037] See also Figure 1 As shown, the present invention is a nearshore sea surface temperature fusion method based on deep learning driven variational analysis, which includes the following steps:

[0038] First, the sea surface temperature data from the previous variational analysis was fed into a deep learning sea surface temperature prediction model. This model extracts spatiotemporal features through three-dimensional convolutional layers, incorporates long-short-term memory units to capture temporal evolution patterns, and employs a multi-step prediction mechanism to simultaneously output multi-day forecasts. The first-day forecast serves as the background for the variational analysis. The model training phase utilizes fused historical satellite data, and the loss function simultaneously constrains spatial gradients and temporal continuity, generating a dynamic, high-quality background field.

[0039] Next, a background error covariance model was constructed based on the predicted background field. By running a deep learning sea surface temperature prediction model, multiple sets of predicted background fields for different forecast periods were generated. The sum of squared differences between each predicted field and the mean field was calculated, and after time series statistics, it was mapped into a spatial error variance field. Furthermore, a fixed-scale spatial Gaussian correlation function and a multi-day sliding window Pearson correlation coefficient were integrated to form a spatiotemporal adaptive correlation structure. This was ultimately combined into a complete background error covariance model that adapts to the characteristics of nearshore short-scale thermal variability.

[0040] Current multi-source sea surface temperature observations are then acquired in real time. Satellite observations include infrared and microwave sensor data. The infrared data has a resolution of approximately 1 km, covers the visible and thermal infrared bands, and includes a single-point quality indicator, sensor bias, and standard deviation information. The microwave data has a resolution of approximately 25 km. In-situ buoy observations come from drifting buoys and profiling buoys, covering surface to shallow water temperatures. After data acquisition, hierarchical quality control and multi-stage bias correction are performed to enhance data reliability.

[0041] The bias-corrected observation data is then assimilated into the predicted background field. By constructing an incremental cost function, with the difference between the analysis field and the background field as the variable, the background error term is constrained using the background error covariance model, and the observation error term is taken from the sensor's native standard deviation. The optimal analysis field is then solved through iterative optimization.

[0042] Finally, the grid resolution of the analysis field is adjusted to the sub-kilometer level, and the output is a nearshore sea surface temperature analysis field covering the coastline to the edge of the continental shelf, containing temperature values ​​and error estimation fields, providing high-precision data for nearshore research.

[0043] In a preferred embodiment of the present invention, the sea surface temperature observation data specifically includes:

[0044] Satellite observation data consists of two types of sensor outputs. Infrared sensor data, covering visible and thermal infrared wavelengths, captures the fine thermal structure of the ocean surface. This data has a spatial resolution of approximately 1 kilometer, and each data point is accompanied by a point quality indicator, sensor bias, and standard deviation information. The point quality indicator quantifies the data's reliability, the sensor bias reflects the systematic error of that device, and the standard deviation characterizes the dispersion of the observations. These metadata provide key information for subsequent data quality assessment and correction. Microwave sensor data has a spatial resolution of approximately 25 kilometers. While lower than infrared data, the microwave band's penetrating properties allow for all-weather observations in complex weather conditions, such as cloud cover, complementing infrared data. Together, these two types of satellite data form a multi-source observation system that balances resolution and continuity. In situ buoy observation data comes from drifting buoys and profiling buoys. Drifting buoys, anchored or moving with ocean currents, transmit real-time surface water temperatures, while profiling buoys dive to varying depths in shallow waters to obtain vertical temperature profiles. These in-situ observation data have been rigorously calibrated and have high measurement accuracy. They can provide ground truth reference for satellite data and effectively improve the reliability of fusion analysis.

[0045] In a preferred embodiment of this invention, the hierarchical quality control specifically includes:

[0046] First, a multi-level screening mechanism was established for satellite observation data. Initial screening was performed based on the quality indicators attached to the data. These indicators are typically assessed by the data producer based on a comprehensive assessment of observation conditions, sensor status, and other parameters. Only the data points with the highest confidence were retained, ensuring the fundamental quality of the input data. For microwave sensor data, the observation time was determined by calculating the solar zenith angle (SZA). The SZA is the angle between the sun's rays and the zenith of the observation point. When this angle exceeds a certain threshold, it is considered a nighttime observation. Daytime observations were excluded because solar radiation during the day creates a warmer layer on the ocean surface, resulting in a significant deviation between the sub-skin temperature measured by microwave sensors and the true sea surface temperature, affecting data accuracy. Furthermore, the standard deviation of the satellite observation data was compared with that of historical sea surface temperature datasets at the same spatial location. Based on statistical principles, if a data point deviates from the historical mean by more than three standard deviations, it is identified as an outlier and removed. This approach filters out observations that significantly deviate from the normal range and prevents outliers from interfering with the analysis results. For on-site buoy data, verify the device type and measurement depth identifier one by one: Different types of buoys have different measurement principles and accuracy, such as drifting buoys and profiling buoys. It is necessary to confirm whether the device type meets the surface temperature observation requirements. The measurement depth identifier is used to determine whether the data comes from the surface water body. Only valid observations from the surface water body are retained to avoid interference with subsequent analysis caused by deep temperature data or equipment failure data.

[0047] Verify the equipment type and measurement depth identification of on-site buoy data and retain valid observation values ​​of surface water bodies.

[0048] In another preferred embodiment of the present invention, the multi-stage deviation correction specifically includes:

[0049] First, nighttime observations are accurately selected using the infrared data's inherent day / night designation or by calculating the solar zenith angle. During the day, the ocean surface is affected by solar radiation, and the temperature difference between the surface and shallow water layers can reach over 3°C. However, at night, the diurnal warming effect is significantly reduced, and the temperature field more closely resembles the true sea surface conditions. Selecting nighttime data can significantly reduce systematic biases caused by the diurnal temperature difference. Second, the sensor's inherent depth bias estimate is deducted. Different sensor types have inherent differences in measured depth. For example, some infrared sensors represent surface water temperature, while some microwave sensors represent shallow water temperature. By deducting the sensor's inherent depth bias estimate, surface or shallow water temperatures are uniformly converted to a standard depth temperature, such as 1 meter. This achieves depth consistency correction for data from different sensors and eliminates bias caused by differences in measured depth. Finally, within a pre-defined spatial grid, such as a 0.25° × 0.25° grid, the average daily deviation of each sensor's data from the in-situ buoy and a designated infrared benchmark is calculated. The Gaussian interpolation method is used to generate a global continuous deviation field from discrete deviation statistics. Gaussian interpolation is based on the principle of spatial distance attenuation and believes that the deviations of adjacent positions are correlated. The closer the distance, the stronger the correlation. This method can smoothly fill the deviation values ​​of unobserved areas to generate a continuous deviation correction field, which is then dynamically subtracted from the original observation values ​​to eliminate systematic deviations between different sensors, ensure that multi-source observation data are fused under the same deviation benchmark, and improve data consistency.

[0050] In another preferred embodiment of the present invention, the deep learning sea surface temperature prediction model specifically includes:

[0051] The model takes as input continuous, gridded three-dimensional sea surface temperature data, encompassing both spatial and temporal dimensions. It processes this input data through a three-dimensional convolutional layer, using convolutional operations to extract the spatiotemporal characteristics of the sea surface temperature field, such as the spatial distribution patterns of fine structures like heat fronts and eddies. Furthermore, the model leverages the time series processing capabilities of long short-term memory (LSTM) units, using a gating mechanism to selectively remember or forget temperature evolution information at historical time steps. This allows it to capture the dynamic patterns of sea surface temperature over time, such as seasonal temperature rises and the movement trajectories of mesoscale eddies.

[0052] The model adopts a multi-step prediction mechanism, which can simultaneously output the sea surface temperature forecast field for multiple future dates, of which the forecast result of the first day is selected as the forecast background field. This multi-step prediction design enables the model to learn the time dependence of the temperature field and improve the prediction ability of the short-term thermal structure evolution. During the model training phase, historical satellite fusion data are used as training samples. These data have undergone preliminary quality control and bias correction and have high reliability. During the training process, the loss function constrains both the spatial gradient and the temporal continuity: the spatial gradient constraint ensures that the spatial variation of the predicted temperature field conforms to the laws of ocean dynamics and avoids unreasonable temperature mutations; the temporal continuity constraint ensures that the temperature evolution of adjacent time steps is smooth, maintaining the temporal consistency of the temperature field, thereby generating a dynamic and high-quality forecast background field.

[0053] In a preferred embodiment of the present invention, the background error covariance model is constructed as follows:

[0054] First, a deep learning sea surface temperature prediction model is run to generate multiple sets of forecast background fields for different forecast periods, such as those covering the next one to seven days. The degree of dispersion of forecast results for different forecast periods is quantified by calculating the sum of squares of the differences between each forecast background field and the average field. This dispersion is then mapped to a spatial error variance field through time series statistical processing, which reflects the distribution of forecast errors at each spatial location.

[0055] To construct the correlation structure, a fixed-scale spatial Gaussian correlation function and a multi-day sliding window Pearson correlation coefficient were combined with equal weights. The spatial Gaussian correlation function, based on the principle of distance decay, describes the weakening of correlations between spatial locations with increasing distance. The multi-day sliding window Pearson correlation coefficient captures short-term temporal correlations in the temperature field by calculating the correlations between sea surface temperature anomalies within different time windows. The weighted superposition of these two factors creates a spatiotemporal adaptive correlation structure that not only reflects the influence of spatial distance on correlations but also adapts to the short-term temporal characteristics of temperature variations in nearshore areas.

[0056] Finally, based on the covariance decomposition principle, the spatial error variance field and the dynamic correlation structure are combined into a complete background error covariance model. This model uses the error variance as diagonal elements and the dynamic correlation structure as off-diagonal elements to construct a covariance matrix that can characterize the spatiotemporal error characteristics of the sea surface temperature field. This provides a scientific basis for the weight allocation of observational data in subsequent variational analysis, enabling the model to better adapt to the complex thermal variability characteristics of nearshore areas.

[0057] In another preferred embodiment of the present invention, the process of solving the optimal analysis field is:

[0058] The optimal analysis field is solved based on a variational analysis framework. An incremental cost function is constructed to fuse the observed data with the predicted background field. This cost function uses the difference between the analysis field and the predicted background field as the optimization variable, aiming to balance the dynamic information of the background field with the observed constraints. The background error term is constrained by a previously constructed background error covariance model. This model characterizes the error distribution of the background field through a spatiotemporally adaptive correlation structure, reflecting the short-scale characteristics of thermal variability in nearshore areas. The observation error term is constructed as a diagonal matrix whose elements are directly derived from the native standard deviation parameters of each sensor. These parameters are pre-calibrated by the sensor data producer based on observation conditions and device characteristics to quantify the uncertainty of single-point observations. The cost function is minimized through an iterative optimization algorithm, and the analysis increments are gradually adjusted. Ultimately, the optimized increments are superimposed on the predicted background field to generate the optimal analysis field that integrates multi-source information. This process essentially seeks the optimal solution between the dynamic prior provided by the background field and the immediate constraints of the observations, ensuring that the analysis field both conforms to historical evolution and responds to the latest observations.

[0059] In a preferred embodiment of this invention, the optimal analysis field generation mechanism demonstrates significant adaptability in the special scenario of missing satellite observation data. In this case, the analysis process relies entirely on the predicted background field generated by the deep learning sea surface temperature prediction model, and local spatial constraints are enhanced through the short-term correlation component in the background error covariance model. Specifically, the short-term correlation component is constructed based on the Pearson correlation coefficient of a multi-day sliding window, which can capture the spatial correlation characteristics of the temperature field in the nearshore area on a short time scale. When observations are missing, this component dynamically adjusts the spatial correlation length scale to strengthen the temperature correlation constraints between adjacent grid points, avoiding spatial discontinuities caused by missing data. Furthermore, the deep learning sea surface temperature prediction model learns the physical laws of the spatiotemporal evolution of the sea surface temperature field during training, such as implicit dynamic characteristics such as heat conduction and fluid motion. The generated predicted background field can maintain the spatiotemporal continuity of the temperature field. For example, when a region has no observations for multiple consecutive days, the model infers a temperature evolution that conforms to physical laws based on historical temperature gradients and time series trends, thereby suppressing the cumulative effect of errors. This dual mechanism ensures the rationality of the analysis field in areas where observations are missing, avoids unreasonable temperature mutations or outliers, and maintains the spatiotemporal consistency of the data field.

[0060] In another preferred embodiment of the present invention, the output of nearshore sea surface temperature analysis utilizes a gridded, regular resampling strategy to achieve high-precision spatial representation. The output strictly covers the critical nearshore region from the coastline to the continental shelf edge, where complex terrain and intensive human activity place an urgent need for high-resolution SST data. The grid resolution is set to sub-kilometer, such as a 0.01° grid. Compared to traditional products with resolutions above 1 kilometer, this allows for the capture of fine-scale thermal structures such as nearshore heat fronts and coral reef microenvironments. The output data utilizes a standardized format and consists of two core components: the sea surface temperature value, which directly reflects the thermal distribution of the nearshore area; and the corresponding error estimate field, calculated based on a background error covariance model and observation error weights, which quantifies the uncertainty of the temperature value at each grid point. The inclusion of the error estimate field enables users to intuitively assess data quality. In areas with sparse observations or complex terrain, the error estimate can be used to adjust the degree of reliance on analysis results. Furthermore, the output data maintains temporal continuity, enabling users to conduct spatiotemporal series analysis and providing high-precision and reliable basic data support for applications such as nearshore ecological monitoring and fishery resource management. The entire output process uses standardized data format and precision control to ensure that the analysis results can directly serve nearshore oceanographic research and business application scenarios.

[0061] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A nearshore sea surface temperature fusion method based on deep learning driven variational analysis, characterized by: The following steps are involved: The historical sea surface temperature data after variational analysis is input into the deep learning sea surface temperature prediction model, and the predicted sea surface temperature for a specified period in the future is output as the prediction background field for the current variational analysis; Based on the deep learning sea surface temperature prediction model, the relationship between the multi-step sea surface temperature prediction difference and the prediction error is established to generate the background error variance during fusion. The background error covariance model for variational analysis is constructed by fusing the spatial distance function and the short-term temporal correlation function. Collect real-time sea surface temperature observation data, including satellite observation data and in-situ buoy observation data, and perform hierarchical quality control and multi-stage bias correction on the sea surface temperature observation data; Assimilating the collected sea surface temperature observation data into the prediction background field, combining the background error covariance model and the observation error weight, and solving the optimal analysis field through variational analysis; Adjusting the grid resolution of the optimal analysis field to sub-kilometer level, and outputting the nearshore sea surface temperature analysis field; The hierarchical quality control specifically includes: The highest confidence data points were selected based on the quality marks attached to the satellite observation data. The solar zenith angle was calculated for the microwave sensor data, and daytime observation data were excluded. The standard deviations of the satellite observation data and the historical sea surface temperature dataset at the same spatial location were compared, and abnormal data points that deviated by more than a set multiple of the standard deviation were eliminated. The equipment type and measurement depth identification of the field buoy data were verified, and the valid observation values ​​of the surface water were retained. The multi-stage deviation correction specifically includes: First, nighttime observation data are filtered using infrared data day / night identification or solar zenith angle calculations. Next, the sensor's native depth bias estimate is deducted to convert surface or shallow water temperatures to standard depth temperatures. Finally, the average daily deviation of each sensor from the on-site buoy and designated infrared benchmark is calculated within a preset spatial grid. A global deviation field is generated through Gaussian interpolation and dynamically deducted from the original observations. The construction process of the background error covariance model is: Run the deep learning sea surface temperature prediction model to generate multiple sets of forecast background fields for different forecast periods. Calculate the sum of squares of the differences between each forecast background field and the forecast mean field, and then map it to a spatial error variance field after time series statistics. The spatial Gaussian correlation function of fixed scale and the Pearson correlation coefficient of multi-day sliding window are integrated with equal weight coefficients, and the weighted superposition is used to form a spatiotemporal adaptive correlation structure. The error variance field and the dynamic correlation structure are combined into a complete background error covariance model according to the covariance decomposition principle.

2. The nearshore sea surface temperature fusion method based on deep learning driven variational analysis according to claim 1 is characterized in that: The sea surface temperature observation data specifically includes: The satellite observation data includes infrared sensor data and microwave sensor data. The infrared sensor data covers visible light and thermal infrared bands. The satellite observation data is accompanied by single-point quality marks, sensor deviation and standard deviation information; the on-site buoy observation data comes from drifting buoys and profile buoys, covering surface to shallow water temperatures.

3. The nearshore sea surface temperature fusion method based on deep learning driven variational analysis according to claim 1 is characterized in that: The deep learning sea surface temperature prediction model specifically includes: The input is continuous gridded 3D sea surface temperature data, and the spatiotemporal features are extracted through a 3D convolutional layer, combined with long short-term memory units to capture the temporal evolution law; A multi-step prediction mechanism is used to synchronously output the forecast fields for multiple days in the future, where the forecast results for the first day are used as the forecast background field for variational analysis. Historical satellite fusion data is used in the model training phase, and the loss function constrains both spatial gradient and temporal continuity.

4. The nearshore sea surface temperature fusion method based on deep learning driven variational analysis according to claim 1 is characterized in that: The solution process of the optimal analysis field is: An incremental cost function is constructed, and the difference between the analysis field and the predicted background field is defined as a variable. The background error term is constrained by the background error covariance model, and the observation error term is set as a diagonal matrix with element values ​​taken from the native standard deviation of the sensor. The cost function is minimized through an iterative optimization algorithm to obtain the analysis increment and superimpose it on the predicted background field to generate the optimal analysis field.

5. The nearshore sea surface temperature fusion method based on deep learning driven variational analysis according to claim 4 is characterized in that: When satellite observation data is missing, the optimal analysis field completely relies on the predicted background field, and the local spatial constraints are enhanced by the short-time correlation components in the background error covariance model; the physical laws implicit in the deep learning sea surface temperature prediction model are used to maintain the spatiotemporal continuity of the temperature field.

6. The nearshore sea surface temperature fusion method based on deep learning driven variational analysis according to claim 1 is characterized in that: The output nearshore sea surface temperature analysis adopts gridding rule resampling, covering the spatial range from the coastline to the edge of the continental shelf, and the grid resolution is set to sub-kilometer level; the output data includes the sea surface temperature value and the corresponding error estimation field.

Images