A method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning

Abstract

The present invention discloses a method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning, which belongs to the technical field of ocean data processing. The method specifically comprises the following steps: firstly, obtaining original satellite sea surface temperature data of a target sea area and preprocessing the data, discretizing the data into a set of dynamic grid cells, wherein each grid contains a spatial position and a sea surface temperature state attribute; then, identifying source dynamic grid cells with valid observation values ​​and target dynamic grid cells with missing data; constructing a spatiotemporal propagation chain for each target grid, which contains a temperature information transmission path and a historical temperature sequence; based on the historical data of the propagation chain and in combination with the physical constraint rules of ocean fluid dynamics, iteratively calculating the current temperature estimate of the target grid; and finally, outputting complete sea surface temperature reconstruction data covering the entire sea area. The present invention effectively solves the problem of missing satellite sea surface temperature data by fusing spatiotemporal features with physical constraints, thereby improving the physical rationality and integrity of the temperature reconstruction data.

Description

Technical Field

[0001] The present invention relates to the technical field of ocean data processing, and in particular to a method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning. Background Art

[0002] As a key parameter characterizing the dynamic and thermal state of the ocean, sea surface temperature (SST) plays an irreplaceable role in climate monitoring, weather forecasting, fishery resource management, and marine environmental research. Satellite remote sensing, with its advantages of large-scale and high-frequency observations, has become the primary means of obtaining global SST data. Infrared sensors can provide high-spatial-resolution data but are susceptible to cloud cover. Microwave sensors, while capable of partially penetrating clouds, have low spatial resolution, and near-shore data are susceptible to land contamination. As a result, single-source satellite SST products generally suffer from discontinuous spatial coverage and severe data loss, a particularly prominent problem in the cloudy and rainy South China Sea. To obtain high-quality SST fields that are continuous in both time and space, it is urgent to develop effective data reconstruction and fusion technologies that fully utilize multi-source satellite observation information.

[0003] To address the issue of missing satellite sea surface temperature data, existing technologies primarily employ two approaches: one is the optimal interpolation method based on statistical interpolation and its improved methods, which interpolate missing areas by constructing a background field covariance matrix and combining it with observation error estimates. The other is a data assimilation-based approach that assimilates satellite observations into numerical forecast models and utilizes the model's physical constraints to generate a complete sea surface temperature field. In recent years, with the advancement of deep learning technology, methods have emerged that use convolutional neural networks or recurrent neural networks to learn the spatiotemporal variations in sea surface temperature and predict temperature values ​​in missing areas. These techniques have, to a certain extent, improved the completeness of data coverage.

[0004] However, the statistical interpolation-based method is highly dependent on the preset background field covariance structure, and it is difficult to accurately characterize the multi-scale ocean dynamic processes in complex sea areas, such as the impact of mesoscale eddies, fronts, and coastal currents on the complex spatiotemporal variability of the temperature field, which can easily lead to deviations in the physical continuity of the reconstruction results. Although the existing deep learning reconstruction methods can automatically learn data features, their network structure design usually lacks explicit modeling and constraints on the inherent physical laws of ocean fluid dynamics, such as heat transport dominated by advection, the reasonable range of temperature gradient changes, and the asymmetric influence of ocean current direction on heat transfer. As a result, the reconstruction results may have physically unreasonable temperature structures, especially in areas where long time series are missing or there are large areas without observations, and their reliability and robustness are significantly reduced. Summary of the Invention

[0005] The purpose of this invention is to provide a method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning to solve the following technical problems:

[0006] The statistical interpolation-based method relies on the preset background field covariance structure, which makes it difficult to characterize the impact of multi-scale ocean dynamic processes in complex sea areas on the temperature field, resulting in physical continuity deviations in the reconstruction results. The existing deep learning reconstruction method lacks explicit modeling and constraints on the physical laws of ocean fluid dynamics, and its reliability and robustness are significantly reduced in areas with long time series missing or large-scale unobserved areas.

[0007] The purpose of the present invention can be achieved through the following technical solutions:

[0008] A method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning includes the following steps:

[0009] S1. Obtain the original satellite sea surface temperature observation data of the target sea area and preprocess it, discretize the surface of the target sea area into a set of dynamic grid cells, each of which has a spatial location attribute and a sea surface temperature state attribute;

[0010] S2. Identifying, based on the original satellite sea surface temperature observation data, source dynamic grid cells with valid observation values ​​and target dynamic grid cells with missing sea surface temperatures in the dynamic grid cell set;

[0011] S3. Constructing a spatiotemporal propagation chain for each target dynamic grid cell, wherein the spatiotemporal propagation chain represents a temperature information transmission path from an adjacent source dynamic grid cell or a reconstructed target dynamic grid cell, and the spatiotemporal propagation chain includes a historical sea surface temperature state sequence of each dynamic grid cell on the transmission path;

[0012] S4. Iteratively calculating the estimated sea surface temperature of each target dynamic grid cell at the current moment based on the historical sea surface temperature state sequence of the dynamic grid cells linked to the spatiotemporal propagation chain and in combination with preset ocean fluid dynamics physical constraints;

[0013] S5. After completing the calculation for all target dynamic grid cells, output complete sea surface temperature reconstruction data covering the entire target sea area.

[0014] As a further solution of the present invention: in S1, the process of preprocessing the original satellite sea surface temperature observation data is:

[0015] Perform basic quality control on the raw satellite sea surface temperature observation data, determine the range defined by the upper and lower limits based on the historical climatological temperature statistics of the target sea area, and remove observations with values ​​outside this range;

[0016] Observation data from different satellite platforms or sensors are bias-matched and adjusted to eliminate systematic offsets between different data sources. The bias-matching adjustment is completed based on the contemporaneous cross-comparison between satellite observation data and matching with field-measured anchor data.

[0017] As a further solution of the present invention: in S1, the process of discretizing the target sea area surface into a set of dynamic grid cells is:

[0018] Adjust the boundary shape and size of the grid cells based on the ocean current direction, monsoon path, and coastline profile of the target sea area. If there is a dominant ocean current direction in the target sea area, align the long axis of the grid cells with the dominant ocean current direction. If there is no dominant ocean current direction in the target sea area but there is a dominant monsoon wind direction, align the long axis of the grid cells with the dominant monsoon wind direction.

[0019] A first grid and a second grid are set, and the size of the second grid is larger than that of the first grid. The first grid is used to divide the sea area within a preset annular distance threshold around the closed center point of the ocean circulation; the second grid is used to divide the sea area where the distance from the coastline is greater than the preset coast distance threshold and the water depth is greater than the preset water depth threshold;

[0020] The spatial location attributes of each dynamic grid cell are defined by the latitude and longitude coordinates of the geometric center point, and the sea surface temperature state attributes are represented by the effective statistics of the original satellite sea surface temperature observations in the cell area; if the cell area is completely in the temperature missing area, its sea surface temperature state attributes are marked as to be reconstructed.

[0021] As a further solution of the present invention: in S3, the process of constructing a spatiotemporal propagation chain for each target dynamic grid unit is:

[0022] With the target dynamic grid cell as the center, all dynamic grid cells located within the first spatial distance threshold are searched in the spatial dimension; if the searched dynamic grid cell has a valid sea surface temperature state record within a preset time length threshold, it is selected as a candidate source dynamic grid cell;

[0023] Analyzing the historical sea surface temperature state change trend of the candidate source dynamic grid cell and its relative position relationship with the target dynamic grid cell;

[0024] Combined with the statistical information of the direction and velocity of the dominant ocean currents during the same period in history, candidate source dynamic grid cells with historical temperature transfer associations with the target dynamic grid cells are screened out;

[0025] Establish a one-way propagation link from each screened candidate source dynamic grid cell to the target dynamic grid cell; each propagation link records the sea surface temperature state value of the candidate source dynamic grid cell at each moment within a preset time length threshold and its temporal and spatial distance relationship with the target dynamic grid cell;

[0026] All candidate propagation links pointing to the same target dynamic grid cell are integrated to form a spatiotemporal propagation chain of the target dynamic grid cell. The spatiotemporal propagation chain consists of multiple data branches, each of which records a historical temperature transmission path and the corresponding historical sea surface temperature state sequence.

[0027] As a further solution of the present invention: in S4, the process of iteratively calculating the estimated sea surface temperature of each target dynamic grid cell at the current moment is:

[0028] For each data branch in the spatiotemporal propagation chain, the gradient variation characteristics of the recorded historical sea surface temperature state sequence in the propagation direction are analyzed;

[0029] According to the gradient variation characteristics and the spatiotemporal distance experienced by the propagation link, the temperature impact component of the target dynamic grid unit at the current moment in the propagation direction of the data branch is calculated;

[0030] The temperature influence components inferred from all data branches in the spatiotemporal propagation chain are integrated to obtain the initial sea surface temperature estimate of the target dynamic grid cell through weighted aggregation.

[0031] The weight factor of the weighted aggregation is generated by the historical stability index of each propagation link, the spatiotemporal proximity index to the target dynamic grid unit, and the consistency index between its propagation direction and the current dominant physical driving direction;

[0032] In each iteration, the most recently reconstructed target dynamic grid cell is regarded as a new candidate source dynamic grid cell, and the spatiotemporal propagation chain of its neighboring target dynamic grid cells that have not yet been reconstructed is updated; the above extrapolation process is repeated until all target dynamic grid cells obtain sea surface temperature estimates.

[0033] As a further solution of the present invention: in S4, the preset ocean fluid dynamics physical constraint rules have the following functions:

[0034] The reconstructed temperature gradient value obtained by calculation is limited to not exceeding the historical maximum temperature gradient change rate determined based on historical statistics of the target sea area;

[0035] If the propagation direction of the data branch is consistent with the dominant ocean current direction at the current moment, a weight factor higher than the average level is assigned to the data branch; if the propagation direction of the data branch is consistent with the dominant wind stress direction at the current moment, a weight factor higher than the average level is assigned to the data branch;

[0036] During weighted aggregation, a first weight factor is applied to the data branch from the downstream direction; a second weight factor is applied to the data branch from the upstream direction; wherein the first weight factor is greater than the second weight factor;

[0037] When the calculated result causes the reconstructed temperature gradient value of the local area to be greater than the historical maximum temperature gradient change rate, the estimated sea surface temperature value of the target dynamic grid cell in the local area is recalculated.

[0038] As a further solution of the present invention: in S4, the reconstruction order of the iterative calculation process is:

[0039] First, all target dynamic grid cells located at the edge of the continuous missing region whose area is larger than the preset area threshold are reconstructed, and the number of candidate source dynamic grid cells contained in the spatiotemporal propagation chain of the target dynamic grid cell must meet the preset valid source number threshold;

[0040] Secondly, reconstructing all target dynamic grid cells within the continuous missing area, and the spatiotemporal propagation chain of the target dynamic grid cell needs to be linked to at least one reconstructed dynamic grid cell. If it is not linked to the reconstructed dynamic grid cell, it needs to be linked to a candidate source dynamic grid cell located within a second spatial distance threshold; wherein the second spatial distance threshold is less than the first spatial distance threshold;

[0041] Finally, the target dynamic grid cell whose number of candidate source dynamic grid cells in the spatiotemporal propagation chain is less than the preset effective source number threshold is processed, and the spatiotemporal propagation chain of the target dynamic grid cell cannot be linked to any reconstructed dynamic grid cell;

[0042] The reconstruction order is to advance from the area where the density of candidate source dynamic grid units is higher than the preset density threshold to the area where the density of candidate source dynamic grid units is lower than the preset density threshold.

[0043] As a further solution of the present invention: in S5, the estimated sea surface temperature of each dynamic grid cell in the complete sea surface temperature reconstruction data is accompanied by a credibility index, and the value range of the credibility index is 0 to 1;

[0044] The credibility index is calculated through normalized weighted linear combination based on the proportion of valid data branches contained in the spatiotemporal propagation chain of the target dynamic grid unit, the satisfaction score of the physical constraint rules during the inference process, and the priority coefficient of the target dynamic grid unit in the reconstruction order.

[0045] As a further solution of the present invention: in said S5, after outputting the complete sea surface temperature reconstruction data, it also includes performing a global consistency check on the reconstructed data:

[0046] Identify a target dynamic grid cell whose reconstructed temperature gradient value is greater than a historical maximum temperature gradient change rate, locally adjust an application threshold of the historical maximum temperature gradient change rate in the physical constraint rule, and if the target dynamic grid cell still cannot meet the setting after the local adjustment, reacquire background information of the area surrounding the target dynamic grid cell;

[0047] A limited number of local iterative optimizations are performed on the target dynamic grid cell until the reconstructed temperature gradient value is less than or equal to the historical maximum temperature gradient change rate.

[0048] Beneficial effects of the present invention:

[0049] The present invention effectively matches complex ocean currents and monsoon environments through dynamic grid unit division technology, overcoming the defect that traditional fixed grids cannot adapt to the evolution of physical processes; the temperature transfer path modeling technology based on the spatiotemporal propagation chain replaces the preset covariance structure in statistical interpolation, and accurately depicts the non-uniform heat transport characteristics driven by dynamic processes such as mesoscale eddies and fronts by mining historical temperature transfer paths and their state sequences; in the iterative calculation process, it deeply integrates the physical constraints of ocean fluid dynamics, including the historical maximum temperature gradient change rate limit, the ocean current direction weight distribution mechanism and the asymmetric advection impact simulation, essentially avoiding physically unreasonable reconstruction results; combined with the reconstruction sequence control strategy that advances from information-rich areas to sparse areas, as well as the global consistency check and backtracking adjustment mechanism, it significantly improves the reconstruction stability of large-scale continuous missing areas; the final output sea surface temperature reconstruction data is accompanied by credibility indicators based on the amount of propagation chain information, physical constraint satisfaction and reconstruction priority, providing a sea surface temperature reconstruction solution for highly dynamic sea areas with both physical consistency, spatiotemporal continuity and reliability quantification. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The present invention will be further described below with reference to the accompanying drawings.

[0051] Figure 1 It is a schematic flow diagram of the present invention. DETAILED DESCRIPTION

[0052] 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.

[0053] See also Figure 1 As shown, the present invention is a method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning, comprising the following steps:

[0054] S1. Obtain the original satellite sea surface temperature observation data of the target sea area and preprocess it. The preprocessing includes basic quality control, determining a reasonable range based on the historical climate temperature statistics of the sea area, eliminating abnormal observations that exceed the range, and adjusting the deviations of the observation data of different satellite platforms or sensors to eliminate systematic offsets. Discrete the surface of the target sea area into a set of dynamic grid cells. When dividing, refer to the direction of ocean currents, monsoon paths and coastline contours to keep the long axis direction of the grid cells consistent with the direction of the dominant ocean currents or monsoons. Smaller grids are used in areas with dense ocean circulation, and larger grids are used in open sea areas. The spatial location attributes of each dynamic grid cell are defined by the longitude and latitude coordinates of the geometric center point, and the sea surface temperature state attributes are characterized by the statistical results of the valid observations in the cell. If it is completely in a missing area, it is marked as a state to be reconstructed.

[0055] S2. Based on the raw satellite sea surface temperature observation data, identify the source dynamic grid cells and the target dynamic grid cells in the dynamic grid cell set. The source dynamic grid cells are grids with valid observation values, such as grids with clear satellite observations in areas without cloud cover. The target dynamic grid cells are grids with missing sea surface temperature data, such as grids in areas covered by clouds or where sensor signals are interrupted.

[0056] S3. Construct a spatiotemporal propagation chain for each target dynamic grid cell. This chain is the path for temperature information to be transmitted. This information can come from adjacent source dynamic grid cells or from reconstructed target dynamic grid cells. The chain contains a sequence of historical sea surface temperature states for each dynamic grid cell along the transmission path, i.e., temperature records at multiple moments in the past. This chain demonstrates the temporal and spatial transmission patterns of temperature, such as how the temperature of a grid cell is transmitted from its neighboring grid cells over time.

[0057] S4. Based on the historical sea surface temperature state sequence of the dynamic grid cells linked by the spatiotemporal propagation chain, combined with the preset ocean fluid dynamics physical constraint rules, iteratively calculate the estimated sea surface temperature of each target dynamic grid cell at the current moment. The constraint rules include limiting the temperature gradient change to no more than the historical maximum range, and the temperature transfer weight in the downstream direction is higher than that in the upstream direction. When calculating, first analyze the temperature gradient change characteristics of each propagation path, combine the spatiotemporal distance to derive the temperature impact component of each path on the target grid, and then obtain the initial estimate through weighted aggregation. In each iteration, the reconstructed target grid will become a new source of information, update the propagation chain of the adjacent unreconstructed grid, and repeat the calculation until all target grids have obtained estimated values.

[0058] S5. After completing the extrapolation for all target dynamic grid cells, the estimated sea surface temperature values ​​of all grids are integrated to output complete sea surface temperature reconstruction data covering the entire target sea area. The estimated value of each grid is also accompanied by a credibility indicator, ultimately forming a continuous sea surface temperature field that conforms to the laws of ocean physics.

[0059] In S1, the process of preprocessing the original satellite sea surface temperature observation data is as follows:

[0060] In the basic quality control phase, we first collect historical climatological temperature statistics accumulated over the years in the target sea area. These data cover the normal temperature fluctuation ranges in different seasons and different sea areas. Based on these historical statistical results, we determine the reasonable range of the current observation data, that is, the upper and lower limits. For example, in a certain sea area in the South China Sea, historical data show that the sea surface temperature in summer mostly fluctuates between 28 and 30 degrees Celsius. If a single point data of 25 degrees Celsius or 33 degrees Celsius appears in a certain satellite observation, and the temperature of the surrounding area of ​​the data is within the normal range, it is determined to be an outlier and removed. For outliers in continuous areas, such as a high temperature area caused by a transient sensor failure or abnormal temperature due to cloud cover, it is necessary to verify it in combination with the observation trend of the adjacent grids. After confirming that it deviates from the overall distribution, it is removed to avoid accidentally deleting extreme values ​​that may reflect real ocean phenomena.

[0061] Multi-source data bias matching and adjustment primarily addresses observational discrepancies between different satellite platforms or sensors. Due to differing detection principles and calibration standards, infrared or microwave sensors carried by different satellites may exhibit systematic offsets in their observations of the same ocean area at the same time. To process these offsets, a cloudless observation period is selected for the same period. The observational data from different satellites are spatially grid-aligned, and temperature differences within the same grid cells are compared. The offset of each satellite's data relative to the reference satellite is recorded. Simultaneously, field data from anchored buoys and ships within the target ocean area is retrieved, and satellite observations are compared with the measured values ​​to correct for inherent sensor errors. For example, nearshore observations from a microwave sensor were consistently 0.4 degrees Celsius lower than the measured values. This was achieved by batch-adjusting all nearshore data from that sensor to align with the measured values. For observational discrepancies between infrared and microwave sensors, microwave data was used as the reference in cloudy areas, and infrared data was used as the reference in clear sky areas. Linear adjustments were performed to eliminate these systematic offsets, ensuring consistent dimensionality and accuracy before fusion of the multi-source data.

[0062] In S1, the process of discretizing the target sea area surface into a set of dynamic grid cells is as follows:

[0063] The boundary shape and size adjustment of the grid unit must be closely combined with the natural dynamic factors of the target sea area. If there is a significant dominant ocean current in the target sea area, such as the Kuroshio branch in the northern South China Sea, the long axis of the grid unit must be consistent with the direction of the ocean current, so that a single grid can fully cover the temperature changes along a section of the ocean current path; if the sea area has no obvious ocean current but is dominated by the monsoon, such as the southwest monsoon in the South China Sea in summer, the long axis of the grid extends along the monsoon wind direction, which is convenient for capturing the horizontal temperature diffusion characteristics driven by wind stress. For nearshore areas with tortuous coastlines, such as the Beibu Gulf coast, the boundaries of the grid units must fit the coastline contour to avoid right-angle grids cutting capes, bays and other terrains, to ensure that the impact of complex nearshore terrain on temperature can be fully represented by the grid units.

[0064] The grid type selection depends on the dynamic activity of the ocean. The first grid size is smaller and is suitable for areas within a preset annular distance threshold around the closed center of the ocean circulation. For example, around the closed center of an anticyclonic vortex in the northern South China Sea, where the vortex's internal waters rotate violently, temperature gradients vary significantly, and multi-scale vortices are nested, a 1 km x 1 km grid size can be used. This small scale accurately captures temperature differences between the vortex's center and its edges. For example, a subtle change of 0.5°C higher temperature at the vortex's center than at its edges can be reflected in the temperature values ​​of adjacent grids, avoiding temperature fluctuations caused by overly large grids that obscure local dynamic processes. The second grid size is larger and is used for open ocean areas with a distance from the coastline greater than a preset coastal distance threshold and a water depth greater than a preset depth threshold. For example, in the central South China Sea, where water depths exceed 2,000 meters and are far from the continental shelf, currents are gentle, the temperature distribution is uniform, and seasonal variations are small, a 2 km x 2 km grid size can be used for the second grid size. Larger grids can cover vast sea areas without increasing unnecessary computational complexity due to over-division, while retaining the overall distribution trend of the temperature field. For example, in summer, the overall temperature in the area remains at around 29°C, and a 2-kilometer-scale grid is sufficient to reflect this large-scale temperature consistency without the need for finer divisions.

[0065] The scale difference between the two stems from different requirements for observation accuracy: the first grid focuses on capturing details in dynamically active areas, while the second grid focuses on efficient coverage of open ocean areas, together forming a dynamic division system that adapts to the spatial heterogeneity of the ocean.

[0066] The attributes of each dynamic grid cell must be clear and traceable. The spatial location attribute is defined by the longitude and latitude coordinates of the geometric center point. For example, the coordinates of the center point of a grid are 15 degrees north latitude and 115 degrees east longitude. This coordinate can accurately locate the specific location of the grid in the sea area. The sea surface temperature state attribute is determined based on the effective statistics of the original observation value. If there are multiple satellite observation points in the grid, the average or median of these points is taken as the representative temperature value of the grid; if the grid is completely blocked by clouds or in a blind spot of sensor observation, resulting in no valid data, its temperature state attribute is marked as a state to be reconstructed. For example, a grid around an archipelago in the South China Sea that is completely blocked due to continuous rain and cloudy weather needs to be reconstructed to fill in the temperature value in subsequent steps.

[0067] Through such preprocessing and discretization, the reliability of the original observation data is guaranteed, and the dynamic grid units can adapt to the ocean dynamic characteristics, laying the foundation for the subsequent source grid and target grid identification and the construction of the space-time propagation chain.

[0068] In S3, the process of constructing a spatiotemporal propagation chain for each target dynamic grid unit is as follows:

[0069] With the target dynamic grid cell as the center, the first spatial distance threshold range is spatially delineated, and all dynamic grid cells within this range are searched. For example, around a target grid in the South China Sea (where temperature data is missing due to cloud cover), 50 kilometers is set as the first spatial distance threshold, and all grid cells within this range are searched. For each searched grid cell, its observation records within the preset time length threshold are checked. If a grid has continuous valid sea surface temperature records in the past 7 days and there are no obvious data gaps, it will be included in the candidate source dynamic grid cell; if a grid only has one isolated record 3 days ago and is missing for the rest of the time, it will be excluded due to insufficient temporal continuity to ensure that the candidate source can provide reliable historical temperature change trends.

[0070] Analyze the relative positional relationships between the historical sea surface temperature state sequences of the candidate source dynamic grid cells and the target dynamic grid cells. For example, the temperature of candidate source A has steadily increased from 29°C to 30°C over the past seven days, demonstrating a continuous upward trend; the temperature of candidate source B has fluctuated slightly between 28.5°C and 29°C. Also, observe their positional relationships: Candidate source A is located northeast of the target grid, while candidate source B is located southwest, with straight-line distances of 30 kilometers and 45 kilometers, respectively, from the target grid. Through this analysis, we can preliminarily determine the temperature variation patterns of the various candidate sources and their spatial correlation with the target grid.

[0071] By combining the direction and velocity statistics of the dominant ocean currents in the target sea area over the same period, candidate sources with historical temperature transfer relationships with the target dynamic grid cells are selected. For example, the dominant ocean current in a certain area of ​​the South China Sea is northeasterly each summer, and historical data shows that this current drives temperature transfer from northeast to southwest. If candidate source A is located northeast of the target grid and its temperature change trend is consistent with the existing temperature gradient surrounding the target grid (for example, when candidate source A warms, the grids to the west of the target grid also warm synchronously), it is determined to have a historical temperature transfer relationship with the target grid. If candidate source C is located northwest of the target grid and the ocean current in this direction has historically had a weak impact on temperature transfer, it may be excluded due to the lack of dynamic relationship, even if it is closer.

[0072] For each selected candidate source dynamic grid cell, a one-way transmission link is established to the target dynamic grid cell. Each link records key information in detail: the candidate source's sea surface temperature state sequence at each moment within a preset time threshold (for example, the daily temperature value of candidate source A over the past seven days), as well as the temporal and spatial distance relationship between the candidate source and the target grid cell. For example, if the straight-line distance between candidate source A and the target grid cell is 30 kilometers, based on historical ocean current velocities, it would take approximately 24 hours for temperature information to propagate from A to the target grid cell. This information collectively forms the core content of the link, providing a temporal and spatial reference for subsequent temperature estimation.

[0073] All candidate propagation links pointing to the same target dynamic grid cell are integrated to form a spatiotemporal propagation chain for that target grid. For example, the spatiotemporal propagation chain for the South China Sea target grid described above might include three data branches: the first, from candidate source A in the northeast, records its 7-day temperature rise trend and 24-hour transmission lag; the second, from candidate source D in the southeast (40 kilometers away, with historical monsoon-driven temperature transmission), records its 5-day temperature fluctuation data and 36-hour transmission lag; and the third, from the reconstructed target grid B (located west of the target grid, reconstructed two days ago), records its reconstructed temperature change sequence. Each branch corresponds to a historical temperature transmission path, collectively providing a multi-dimensional reference for temperature estimation in the target grid, ensuring the comprehensiveness and relevance of the transmitted information.

[0074] Through this process, the space-time propagation chain not only covers the temperature information of nearby effective observation sources, but also incorporates the transmission laws dominated by ocean dynamic processes, laying a data foundation for the subsequent calculation of the target grid temperature in combination with physical constraints.

[0075] In S4, the process of iteratively calculating the estimated sea surface temperature of each target dynamic grid cell at the current moment is:

[0076] Analyze the historical sea surface temperature state sequence of each data branch in the spatiotemporal propagation chain, focusing on capturing the gradient variation characteristics along the propagation direction. For example, one data branch from a target dynamic grid cell in the South China Sea originates from candidate source A in the northeast. Its historical temperature sequence shows an increase from 28°C to 29.5°C over the past five days. Furthermore, along the propagation path from A to the target grid, the temperature decreases by 0.3°C per day, forming a stable gradient attenuation trend. Another branch originates from candidate source B in the southeast, with temperatures fluctuating between 28.2°C and 28.5°C. The gradient variation along the propagation path is relatively gentle, decreasing by only 0.1°C per day. This analysis can clarify the temperature transmission patterns of different branches, with some showing continuous attenuation and others fluctuating slowly, providing basic characteristics for subsequent inferences.

[0077] Based on the gradient variation characteristics and the spatiotemporal distance of the propagation link, the temperature impact component of each branch on the target dynamic grid cell is estimated. For example, candidate source A is 30 kilometers away from the target grid. Historical ocean current data shows that it takes approximately 24 hours for temperature information to propagate from source A to the target grid. Combined with its daily attenuation gradient of 0.3°C, the temperature impact component of source A on the target grid at the current moment can be estimated to be 29.5°C minus the attenuation value over 24 hours. Candidate source B is 45 kilometers away from the target grid, with a transmission lag of approximately 36 hours. Its gentle gradient variation means that the impact component is closer to source B's current temperature value. This calculation accounts for both the natural temporal variation of temperature and the attenuation effect caused by spatial distance, making the impact component more consistent with actual transmission patterns.

[0078] The temperature influence components of all data branches are combined, and an initial sea surface temperature estimate for the target dynamic grid cell is obtained through weighted aggregation. The generation of weight factors requires the integration of multiple dimensional indicators: the historical stability of the propagation link is an important reference. If the temperature gradient deviation of a branch has been consistently small during multiple transmissions in the past, its weight will be increased accordingly. The temporal and spatial proximity to the target grid is also critical. Branches with closer distances and shorter transmission time lags generally have higher weights. The consistency of the propagation direction with the current dominant physical driving direction also affects the weight. If the branch direction is consistent with the current ocean current direction, the weight will be significantly increased. For example, of the three data branches for a target grid in the South China Sea, branch 1 comes from the direction of the ocean current, has high historical stability and is close to the current, and has the highest weight. Branch 2, although close, has a large angle between the propagation direction and the ocean current, and has the second highest weight. Branch 3 is far away but has medium stability and the lowest weight. Through this weighted aggregation, the initial estimate can integrate the effective information of each branch, taking into account both reliability and relevance.

[0079] The iterative process is the core of the inference process. After each iteration, the target dynamic grid cell that has completed reconstruction will become a new candidate source and participate in the spatiotemporal propagation chain update of the adjacent unreconstructed grid. For example, in a certain area of ​​the South China Sea, the target grid C at the edge is first reconstructed, and its temperature estimate is included in the propagation chain of the surrounding unreconstructed grid D as a new data branch. As a result, a new path from C is added to the propagation chain of D, which records the temperature change sequence of C and the spatiotemporal distance from D, so that the inference of D can be combined with the latest reconstruction results, reducing dependence on the original candidate source. This iteration continues until all target dynamic grid cells obtain temperature estimates, ensuring the spatial consistency of the reconstruction results.

[0080] In S4, the preset ocean fluid dynamics physical constraint rules have the following functions:

[0081] Reconstructed temperature gradients must be constrained to not exceed the historical maximum temperature gradient change rate for the target sea area. For example, historical data from a certain area in the South China Sea show that, influenced by ocean currents and monsoons, the maximum daily temperature gradient change does not exceed 0.5°C / km. If the temperature difference between a target grid and its neighboring grid to the west exceeds the historical maximum by 0.8°C / km, the transmission chain for that grid must be reviewed to check whether key candidate sources have been missed or whether the gradient analysis is inaccurate, to avoid sudden changes that violate the laws of ocean physics.

[0082] The weight factors of data branches are adjusted based on the current dominant ocean current and wind stress directions. If the propagation direction of a data branch is consistent with the current ocean current direction, for example, if the ocean current flows from northeast to southwest, and the branch originates from a candidate source in the northeast, its weight will be higher than the average level, because the ocean current accelerates the transmission of temperature information, making the reference value of this branch higher. If the propagation direction of a branch is consistent with the current dominant wind stress direction, for example, if the branch originates from a candidate source in the southwest driven by the summer southwest monsoon, its weight will also be increased, because wind stress drives the surface water to flow, strengthening the temperature transmission in that direction.

[0083] Different weighting factors are assigned to the downstream and upstream data branches, with the first downstream weighting factor being greater than the second upstream weighting factor. For example, if the current ocean current flows from east to west, candidate source E in the downstream direction is located east of the target grid. Its temperature information is naturally transmitted to the target grid with the ocean current, so the first weighting factor is higher. Candidate source F in the upstream direction is located west of the target grid. Since the temperature information must be transmitted against the ocean current, which is more difficult and less reliable, the second weighting factor is correspondingly lower. This distinction makes the aggregation results more consistent with the heat transport patterns dominated by the ocean current and reduces the interference of noisy data in the upstream direction.

[0084] When the reconstructed temperature gradient value in a local area exceeds the historical maximum temperature gradient change rate, the target dynamic grid cells in that area must be recalculated. For example, due to the complex terrain around a South China Sea archipelago, the temperature difference between two adjacent grid cells in one recalculation reached 0.6°C / km, exceeding the historical maximum of 0.5°C / km. In this case, the spatiotemporal propagation chain between the two grid cells must be re-examined, and candidate nearby sources that may have been overlooked (such as the temperature influence of small-scale circulation between the archipelagoes) must be added. Alternatively, topography-related parameters in the weighting factors can be adjusted to reduce the recalculated temperature gradient to a reasonable range, ensuring that the temperature changes in the local area meet the ocean dynamic constraints.

[0085] Through this iterative extrapolation that combines the space-time propagation chain with physical constraints, the temperature estimate of the target dynamic grid cell can not only inherit the changing trend of historical data, but also conform to the laws of ocean dynamics, ultimately forming a continuous and physically reasonable sea surface temperature field.

[0086] In S4, the reconstruction order of the iterative calculation process is:

[0087] The reconstruction starts from the edge of the continuous missing area, where the area of ​​these areas is larger than the preset area threshold. The target dynamic grid cells at the edge are processed first because they are closer to the source dynamic grid cells with valid observation values, and the number of candidate source dynamic grid cells included in the spatiotemporal propagation chain can meet the preset threshold of the number of valid sources. For example, in a continuous missing area covered by clouds in the South China Sea, the edge grid is close to the unobstructed source grid, and enough candidate source data (such as temperature observations in multiple directions around it) can be collected in the propagation chain. Prioritizing the reconstruction of these edge grids can form a reliable "transition zone", providing a stable reference benchmark for the subsequent calculation of the internal grids, and avoiding the accumulation of errors caused by jumping directly from the source grid to the internal grid.

[0088] Next, all target dynamic grid cells inside the continuous missing area are processed. The spatiotemporal propagation chains of these internal grids need to be linked to at least one reconstructed dynamic grid cell - that is, the edge grid or other inferred internal grid completed previously. If it is temporarily impossible to link to the reconstructed grid, it is necessary to rely on the candidate source dynamic grid cells within the second spatial distance threshold, and the second spatial distance threshold is smaller than the first spatial distance threshold. For example, an internal grid is far away from the edge reconstructed grid, and the propagation chain cannot be directly associated, but there are source grids that are not completely blocked in a smaller range around it, and they can be inferred based on these close-range source data. This setting ensures that the temperature estimate of the internal grid either inherits the reliable results of the reconstructed grid or relies on closer source data to reduce the deviation caused by long-distance transmission.

[0089] The final target is the target dynamic grid cell whose number of candidate source dynamic grid cells in the spatiotemporal propagation chain falls below the preset threshold for valid sources and cannot be linked to any reconstructed dynamic grid cells. These grid cells are often located in areas with extremely sparse source data, such as deep ocean areas far from continents or isolated areas permanently covered by thick cloud. Very few candidate sources can be collected in the propagation chain, perhaps even just one or two scattered observations. Placing them last allows for the reverse complement of propagation chain information for these grid cells after reconstruction is completed in other areas, leveraging the surrounding continuous temperature field. For example, the temperature gradient of adjacent reconstructed grid cells can be used to infer reasonable values ​​for these isolated grid cells, minimizing errors caused by insufficient source data.

[0090] The overall reconstruction order always proceeds from areas where the density of candidate source dynamic grid cells exceeds a preset density threshold, moving toward areas where the density is below the preset density threshold. Areas with a high density of candidate sources indicate more reliable observational data to support the reconstruction. A stable temperature baseline is established in these areas first, and then the reconstruction is gradually expanded to areas with sparse source data. For example, nearshore waters typically have more abundant source data (assisted by land-based observations). Reconstruction is prioritized for nearshore areas, and then based on this data, it is extended to offshore areas. This ensures that the reconstruction of offshore areas relies on the continuous temperature field gradient nearshore, reducing the overall bias caused by insufficient source data.

[0091] In S5, in the complete sea surface temperature reconstruction data, the estimated sea surface temperature of each dynamic grid cell is accompanied by a credibility index, whose value range is between 0 and 1, and the higher the value, the stronger the reliability of the estimated value.

[0092] The credibility index is calculated by combining three core elements through a normalized, weighted linear combination. The first element is the percentage of valid data branches included in the spatiotemporal propagation chain for the target dynamic grid cell. For example, if a grid cell's propagation chain has five data branches, three of which have continuous historical temperature series that match physical constraints (considered valid), and two are excluded due to data gaps or excessive deviations, the validity percentage is 3 / 5. A higher validity percentage indicates more reliable information supporting the grid cell's temperature estimate, and thus a greater contribution to credibility.

[0093] The second factor is the satisfaction score for the physical constraints during the inference process. These constraints include the range of temperature gradients and the influence of current direction on transport. If the grid temperature estimate fully meets all constraints (e.g., the gradient does not exceed the historical maximum, and the downstream weight distribution is reasonable), the satisfaction score is close to 1. If there are slight deviations (e.g., the gradient slightly exceeds but meets the constraint after local adjustments), the score may drop to 0.7. If multiple adjustments still fail to meet the core constraints, the score will be even lower. The satisfaction score directly reflects the degree to which the estimate conforms to the laws of ocean dynamics and is an important support for its credibility.

[0094] The third factor is the priority coefficient of the target dynamic grid cell in the reconstruction order. Edge grids or grids in areas with rich source data are prioritized in the reconstruction order because they rely on more effective information and have higher priority coefficients (e.g., 0.9). Grids in areas with sparse source data, which are processed later, have lower priority coefficients (e.g., 0.5). This coefficient reflects the grid's position in the overall reconstruction logic. Prioritized grids often provide a benchmark for subsequent reconstructions, providing a more reliable foundation for their estimates.

[0095] These three factors are normalized individually, adjusting them to a range of 0 to 1, and then linearly combined based on their influence on credibility, ultimately yielding a credibility index for each mesh. For example, a prioritized edge mesh with a valid data branch ratio of 0.8, a physical constraint satisfaction score of 0.9, and a priority coefficient of 0.9 might achieve a credibility index of 0.85 after weighted combination. Meanwhile, a later-processed isolated mesh with a valid branch ratio of 0.3, a physical constraint satisfaction score of 0.6, and a priority coefficient of 0.5 might only have a credibility index of 0.4.

[0096] In S5, after outputting the complete sea surface temperature reconstruction data, a global consistency check of the reconstruction data is also performed:

[0097] Before outputting the complete sea surface temperature reconstruction data, a global consistency check is required to ensure that the overall temperature field conforms to the laws of ocean physics.

[0098] First, target dynamic grid cells are identified where the reconstructed temperature gradient exceeds the historical maximum temperature gradient change rate. For example, historical data for a certain area in the South China Sea shows a maximum temperature gradient of 0.5°C / km. If the temperature difference between two adjacent grid cells reaches 0.6°C / km, far exceeding the historical maximum, these grid cells are marked as candidates for optimization.

[0099] For the marked grid, first locally adjust the threshold for the historical maximum temperature gradient change rate in the physical constraint rule. For example, if the marked grid is located in a narrow strait with turbulent ocean currents, where slightly higher gradients have occasionally occurred historically, the threshold for that local area can be temporarily adjusted to 0.55°C / km. If the grid gradient meets the requirements after the adjustment, no further action is required.

[0100] If adjusting the threshold still fails to meet the set requirements, it is necessary to re-acquire background information for the area surrounding the marked dynamic grid cell. For example, if a marked grid cell is located near an archipelago, the temperature influence of the small-scale circulation between the archipelagos may be missed. Therefore, it is necessary to supplement the area with high-resolution satellite observation debris data or in-situ buoy records to improve the information on the spatiotemporal transmission chain and provide a more comprehensive basis for re-estimation.

[0101] Finally, a limited number of local iterative optimizations are performed on the marked target dynamic grid cells. For example, for a marked grid, the weight factors of each branch in the propagation chain are readjusted (e.g., increasing the weight of the downstream branch), or newly identified candidate source data is added to recalculate the temperature estimate. After each optimization, the gradient is checked to ensure it meets the requirements. If the gradient falls within the historical maximum range after three iterations, the optimization is terminated. If the gradient still fails to meet the requirements after a preset number of iterations, the low confidence level of the grid is clearly marked in the output data, along with an explanation of the optimization process for reference in subsequent applications.

[0102] 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 method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning, characterized in that: The following steps are involved: S1. Obtain the original satellite sea surface temperature observation data of the target sea area and preprocess it, discretize the surface of the target sea area into a set of dynamic grid cells, each of which has a spatial location attribute and a sea surface temperature state attribute; S2. Identifying, based on the original satellite sea surface temperature observation data, source dynamic grid cells with valid observation values ​​and target dynamic grid cells with missing sea surface temperatures in the dynamic grid cell set; S3. Constructing a spatiotemporal propagation chain for each target dynamic grid cell, wherein the spatiotemporal propagation chain represents a temperature information transmission path from an adjacent source dynamic grid cell or a reconstructed target dynamic grid cell, and the spatiotemporal propagation chain includes a historical sea surface temperature state sequence of each dynamic grid cell on the transmission path; S4. Iteratively calculating the estimated sea surface temperature of each target dynamic grid cell at the current moment based on the historical sea surface temperature state sequence of the dynamic grid cells linked to the spatiotemporal propagation chain and in combination with preset ocean fluid dynamics physical constraints; S5. After completing the calculation for all target dynamic grid cells, output complete sea surface temperature reconstruction data covering the entire target sea area; In S3, the process of constructing a spatiotemporal propagation chain for each target dynamic grid unit is as follows: With the target dynamic grid cell as the center, all dynamic grid cells located within the first spatial distance threshold are searched in the spatial dimension; if the searched dynamic grid cell has a valid sea surface temperature state record within a preset time length threshold, it is selected as a candidate source dynamic grid cell; Analyzing the historical sea surface temperature state change trend of the candidate source dynamic grid cell and its relative position relationship with the target dynamic grid cell; Combined with the statistical information of the direction and velocity of the dominant ocean currents during the same period in history, candidate source dynamic grid cells with historical temperature transfer associations with the target dynamic grid cells are screened out; Establish a one-way propagation link from each screened candidate source dynamic grid cell to the target dynamic grid cell; each propagation link records the sea surface temperature state value of the candidate source dynamic grid cell at each moment within a preset time length threshold and its temporal and spatial distance relationship with the target dynamic grid cell; All candidate propagation links pointing to the same target dynamic grid cell are integrated to form a spatiotemporal propagation chain of the target dynamic grid cell. The spatiotemporal propagation chain consists of multiple data branches, each of which records a historical temperature transmission path and the corresponding historical sea surface temperature state sequence.

2. The method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning according to claim 1, characterized in that: In S1, the process of preprocessing the original satellite sea surface temperature observation data is as follows: Perform basic quality control on the raw satellite sea surface temperature observation data, determine the range defined by the upper and lower limits based on the historical climatological temperature statistics of the target sea area, and remove observations with values ​​outside this range; Observation data from different satellite platforms or sensors are bias-matched and adjusted to eliminate systematic offsets between different data sources. The bias-matching adjustment is completed based on the contemporaneous cross-comparison between satellite observation data and matching with field-measured anchor data.

3. The method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning according to claim 1, characterized in that: In S1, the process of discretizing the target sea area surface into a set of dynamic grid cells is as follows: Adjust the boundary shape and size of the grid cells based on the ocean current direction, monsoon path, and coastline profile of the target sea area. If there is a dominant ocean current direction in the target sea area, align the long axis of the grid cells with the dominant ocean current direction. If there is no dominant ocean current direction in the target sea area but there is a dominant monsoon wind direction, the long axis direction of the grid cell is aligned with the dominant monsoon wind direction; A first grid and a second grid are set, and the size of the second grid is larger than that of the first grid. The first grid is used to divide the sea area within a preset annular distance threshold around the closed center point of the ocean circulation; the second grid is used to divide the sea area where the distance from the coastline is greater than the preset coast distance threshold and the water depth is greater than the preset water depth threshold; The spatial location attributes of each dynamic grid cell are defined by the latitude and longitude coordinates of the geometric center point, and the sea surface temperature state attributes are represented by the effective statistics of the original satellite sea surface temperature observations in the cell area; if the cell area is completely in the temperature missing area, its sea surface temperature state attributes are marked as to be reconstructed.

4. The method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning according to claim 1, characterized in that: In S4, the process of iteratively calculating the estimated sea surface temperature of each target dynamic grid cell at the current moment is: For each data branch in the spatiotemporal propagation chain, the gradient variation characteristics of the recorded historical sea surface temperature state sequence in the propagation direction are analyzed; According to the gradient variation characteristics and the spatiotemporal distance experienced by the propagation link, the temperature impact component of the target dynamic grid unit at the current moment in the propagation direction of the data branch is calculated; The temperature influence components inferred from all data branches in the spatiotemporal propagation chain are integrated to obtain the initial sea surface temperature estimate of the target dynamic grid cell through weighted aggregation. The weight factor of the weighted aggregation is generated by the historical stability index of each propagation link, the spatiotemporal proximity index to the target dynamic grid unit, and the consistency index between its propagation direction and the current dominant physical driving direction; In each iteration, the most recently reconstructed target dynamic grid cell is regarded as a new candidate source dynamic grid cell, and the spatiotemporal propagation chain of its neighboring target dynamic grid cells that have not yet been reconstructed is updated; the above extrapolation process is repeated until all target dynamic grid cells obtain sea surface temperature estimates.

5. The method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning according to claim 4, characterized in that: In S4, the preset ocean fluid dynamics physical constraint rules have the following functions: The reconstructed temperature gradient value obtained by calculation is limited to not exceeding the historical maximum temperature gradient change rate determined based on historical statistics of the target sea area; If the propagation direction of the data branch is consistent with the dominant ocean current direction at the current moment, a weight factor higher than the average level is assigned to the data branch; if the propagation direction of the data branch is consistent with the dominant wind stress direction at the current moment, a weight factor higher than the average level is assigned to the data branch; During weighted aggregation, a first weight factor is applied to the data branches from the downstream direction; applying a second weighting factor to the data branch from the upstream direction; wherein the first weighting factor is greater than the second weighting factor; When the calculated result causes the reconstructed temperature gradient value of the local area to be greater than the historical maximum temperature gradient change rate, the estimated sea surface temperature value of the target dynamic grid cell in the local area is recalculated.

6. The method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning according to claim 1, characterized in that: In S4, the reconstruction order of the iterative calculation process is: First, all target dynamic grid cells located at the edge of the continuous missing region whose area is larger than the preset area threshold are reconstructed, and the number of candidate source dynamic grid cells contained in the spatiotemporal propagation chain of the target dynamic grid cell must meet the preset valid source number threshold; Secondly, reconstructing all target dynamic grid cells within the continuous missing area, and the spatiotemporal propagation chain of the target dynamic grid cell needs to be linked to at least one reconstructed dynamic grid cell. If it is not linked to the reconstructed dynamic grid cell, it needs to be linked to a candidate source dynamic grid cell located within a second spatial distance threshold; wherein the second spatial distance threshold is less than the first spatial distance threshold; Finally, the target dynamic grid cell whose number of candidate source dynamic grid cells in the spatiotemporal propagation chain is less than the preset effective source number threshold is processed, and the spatiotemporal propagation chain of the target dynamic grid cell cannot be linked to any reconstructed dynamic grid cell; The reconstruction order is to advance from the area where the density of candidate source dynamic grid units is higher than the preset density threshold to the area where the density of candidate source dynamic grid units is lower than the preset density threshold.

7. The method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning according to claim 1, characterized in that: In S5, the estimated sea surface temperature of each dynamic grid cell in the complete sea surface temperature reconstruction data is accompanied by a credibility index, and the value range of the credibility index is 0 to 1; The credibility index is calculated through normalized weighted linear combination based on the proportion of valid data branches contained in the spatiotemporal propagation chain of the target dynamic grid unit, the satisfaction score of the physical constraint rules during the inference process, and the priority coefficient of the target dynamic grid unit in the reconstruction order.

8. The method for reconstructing missing sea surface temperature data based on spatiotemporal feature learning according to claim 1, characterized in that: In S5, after outputting the complete sea surface temperature reconstruction data, a global consistency check of the reconstruction data is also performed: Identify a target dynamic grid cell whose reconstructed temperature gradient value is greater than a historical maximum temperature gradient change rate, locally adjust an application threshold of the historical maximum temperature gradient change rate in the physical constraint rule, and if the target dynamic grid cell still cannot meet the setting after the local adjustment, reacquire background information of the area surrounding the target dynamic grid cell; A limited number of local iterative optimizations are performed on the target dynamic grid cell until the reconstructed temperature gradient value is less than or equal to the historical maximum temperature gradient change rate.

Images