Ocean spatiotemporal data interpolation method based on remote sensing conditional information diffusion

Abstract

This invention relates to the field of marine data processing and intelligent reconstruction technology, and in particular provides a marine spatiotemporal data interpolation method based on remote sensing conditional information diffusion. The method includes processing observation data from multiple ground stations and remote sensing data from L4 ocean to obtain observation datasets, remote sensing datasets, and datasets with different missing rates; standardizing and aligning the observation and remote sensing datasets to obtain preprocessed data and a trained model; constructing an interpolation architecture based on a conditional diffusion mechanism; and using the trained model to perform interpolation inference for datasets with different missing rates. This method improves the accuracy of interpolating missing values ​​in observation data and ensures the physical rationality and overall continuity of the interpolation results.

Description

Technical Field

[0001] This invention relates to the field of marine data processing and intelligent reconstruction technology, and in particular to a marine spatiotemporal data interpolation method based on remote sensing conditional information diffusion. Background Technology

[0002] For a long time, ocean observation data has been of paramount importance in areas such as global climate change monitoring, marine disaster early warning, resource development, and environmental protection. High-quality, continuous ocean data is not only crucial for scientific research but also forms the basis for various engineering decisions. However, due to issues such as the sparse distribution of ocean observation stations, high equipment maintenance costs, and limitations imposed by weather and environmental conditions, ocean data generally suffers from severe spatiotemporal gaps. This data incompleteness significantly weakens the reliability of subsequent tasks such as ocean data analysis, modeling, and forecasting.

[0003] Traditional ocean data interpolation methods mainly include physical model-based methods and statistical methods. Physical model-based methods are often computationally complex and lack flexibility, making them difficult to adapt to complex and ever-changing ocean observation data. Statistical methods, such as Kriging interpolation and nearest neighbor interpolation, often rely on strong specific assumptions, making it difficult to handle nonlinear spatiotemporal relationships in the ocean, resulting in limited interpolation effectiveness.

[0004] In recent years, with the rapid development of artificial intelligence technology, data-driven methods, such as machine learning and deep learning, have been gradually introduced into the processing of marine time-series data, achieving certain results, especially in marine data prediction. For example, deep learning methods such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and gated recurrent units (GRUs) have shown certain advantages in time series prediction. However, most methods focus on time-dimensional dependency modeling, often neglecting the spatial coupling relationship between observation stations. Furthermore, when the missing data ratio is high, errors tend to accumulate during iteration, leading to a decrease in interpolation accuracy. Interpolation, as a fundamental data repair step, is crucial for improving the accuracy of subsequent prediction models, enhancing data availability, and improving overall system stability. Accurate data interpolation not only fills in observational blind spots but also provides a reliable data foundation for long-term series analysis and extreme event detection. Overall, existing technologies do not focus enough on repairing missing data, prioritizing future prediction, resulting in a significant gap in the critical step of restoring the integrity of marine data.

[0005] With the development of observation methods, remote sensing technology has gradually become an important source of information for ocean research. Remote sensing data has the advantages of high spatial resolution and wide coverage, and can reflect key elements of the ocean surface, such as sea surface temperature, sea wind field, and ocean currents, on a large scale. However, remote sensing observations are limited by orbital periods and meteorological conditions, resulting in generally insufficient temporal resolution, with most data being at the day level or even lower frequency. While in-situ observations have advantages in temporal resolution, they are difficult to support continuous and complete spatiotemporal modeling due to limited locations and severe data gaps. Summary of the Invention

[0006] In view of this, the present invention provides a marine spatiotemporal data interpolation method based on remote sensing conditional information diffusion, which improves the accuracy of interpolation of missing values ​​in observation data and ensures the physical rationality and overall continuity of the interpolation results.

[0007] In a first aspect, the present invention provides a method for interpolating marine spatiotemporal data based on remote sensing conditional information diffusion, the method comprising:

[0008] Step 1: Process the observation data from multiple ground stations and the remote sensing data from the L4 ocean to obtain the observation dataset, the remote sensing dataset, and datasets with different missing rates.

[0009] Step 2: Standardize and align the observation dataset and remote sensing dataset respectively to obtain preprocessed data and the trained model;

[0010] Step 3: Construct an imputation architecture based on the conditional diffusion mechanism, and use the trained model to perform imputation inference on datasets with different missing rates.

[0011] Optionally, step 1 includes:

[0012] a. Observation data processing:

[0013] After acquiring the observation data, it was stored in the form of .nc files for multiple stations. First, the data was read and converted to .csv format. It was then uniformly converted into a standardized date and time representation. Second, the observation data consisted of multiple depths and variables, and only the temperature TEMP variable at a depth of 0.5 m for each station was retained. Finally, missing values ​​were processed and set to NaN. Subsequently, the data from all stations were merged to generate the observation data file, i.e., the observation dataset .csv. At the same time, the latitude and longitude information of each station was extracted to generate the latitude and longitude file, i.e., the latitude and longitude .csv. To improve the model's adaptability to different missing rates, missing datasets with different missing rates were manually constructed based on the real observation data.

[0014] b. Remote sensing data processing:

[0015] The remote sensing data for L4 ocean is hourly remote sensing grid data; based on the latitude and longitude information of each station, the data within the corresponding grid is extracted as the station's remote sensing data; since both the observation data and the remote sensing data are hourly data, they need to be aligned in the time dimension; finally, a remote sensing dataset, i.e., remote sensing dataset.csv, is generated.

[0016] Through the above processing, we obtain observation datasets, remote sensing datasets, and datasets with different missing rates.

[0017] Optionally, obtaining the preprocessed data in step 2 includes:

[0018] Observational data standardization: Calculate the mean and standard deviation of the training set and normalize the original data using a standardization formula; Remote sensing data distribution alignment: Map the data to a distribution range similar to the observation data through statistical feature calculations to eliminate scale differences between different data sources.

[0019] When constructing training samples, a time step sequence is generated, and observation data masks and conditional masks are constructed. The observation data, remote sensing data, and corresponding masks are integrated to form the data batch required for training.

[0020] Optionally, the trained model is obtained in step 2, and its training process includes:

[0021] The preprocessed data is input into the interpolation model, forward propagation is performed to obtain the prediction results, and the loss between the predicted value and the true value is calculated. The loss function comprehensively considers the data reconstruction accuracy, remote sensing data constraints, and spatiotemporal consistency. The gradient is calculated through the backpropagation algorithm, and the model parameters are updated using the Adam optimizer. The above training process is iterated until the model converges. Finally, the trained model is saved.

[0022] Optionally, step 3 includes: in the inference phase, firstly, after loading the model, the input observation dataset and remote sensing dataset are standardized and aligned in the same way as in the training phase to generate time series codes and corresponding masks; subsequently, an interpolation framework based on a conditional diffusion mechanism is constructed, the conditional diffusion mechanism of which includes:

[0023] h. During the forward diffusion phase, the model gradually adds Gaussian noise to the original sea surface temperature data, causing it to evolve into a form close to a standard normal distribution. This process constitutes a Markov chain, defined as:

[0024] ;

[0025] ;

[0026] in, Indicates diffusion time step Data; Indicates diffusion time step Data; This represents the preset noise scheduling parameters used to control the intensity of noise injection at each step; I represents the identity matrix. Represents the original data; This represents the data after adding noise at all time steps from 1 to T, i.e., the final noisy data; as the diffusion mechanism progresses, the original data structure is gradually lost, and the synchronously generated time series code and corresponding mask will be used for conditional control in the reverse diffusion stage;

[0027] i. In the inverse diffusion stage, the model starts from a pure Gaussian noise distribution and uses the inverse diffusion algorithm to iteratively denoise the noise; in each step, the model relies on conditional information. ,in Indicates time characteristics, Indicates spatial characteristics, Representing remote sensing features; the target data of the missing area is reconstructed step by step through a denoising process. The inverse diffusion process is defined as:

[0028] ;

[0029] in, This indicates that the noise prediction model is based on the data from the diffusion time step t. Time step t and condition information The predicted mean term; This represents the predefined learnable noise variance;

[0030] j. In the post-processing stage, after completing the inverse diffusion process, the system de-standardizes the prediction results to map them back to the original data distribution range. Then, the system performs spatiotemporal consistency checks and physical rationality verification on the interpolation results to ensure that the prediction data and the original observation data are spatially and temporally consistent. Finally, the system integrates the interpolation results with the original observation data to generate complete and continuous spatiotemporal sequence data for data completion.

[0031] Optionally, the interpolation architecture of the marine remote sensing conditional diffusion mechanism consists of an input layer, a conditional information construction module, a noise prediction module, and an output layer;

[0032] First, in the forward diffusion stage, the model gradually injects noise into the original data, transforming it into noise distribution data. In the reverse diffusion stage, guided by conditional information, the model gradually reconstructs the target data through an iterative denoising process. To model the spatiotemporal characteristics of the ocean, a conditional information construction module is designed to perform triple feature extraction, extracting temporal and spatial features through Transformer and Graph Convolutional Neural Network respectively, and combining remote sensing data to enhance observation information and extract remote sensing features. The noise prediction model internally includes temporal attention, spatial attention, and Graph Neural Network (GNN). After optimization through multiple loss functions, the generated interpolated data is finally combined with the original observation data to form a complete, spatiotemporally continuous ocean dataset.

[0033] Optionally, the triple feature extraction and conditional information construction includes:

[0034] To comprehensively capture the spatiotemporal dynamics of marine environmental data, this study combines temporal feature extraction, spatial feature extraction, and feature encoding based on remote sensing feature extraction. It also incorporates conditional information to guide the inverse diffusion process in the diffusion model. Triple feature extraction is performed through a temporal feature extraction module, a spatial feature extraction module, and a remote sensing feature extraction module. The temporal feature extraction module captures the evolution of marine data at different time scales; the spatial feature extraction module characterizes the spatial correlation and similarity between observation stations; and the remote sensing feature extraction module incorporates large-scale remote sensing information to provide spatial background constraints for local observations.

[0035] I. Temporal Feature Extraction: For the multi-scale temporal features of sea surface temperature series data, a temporal feature extraction module based on an improved Transformer structure is employed. First, the input data is fed into a temporal feature stream, and time points in the time series are mapped to a high-dimensional feature space through temporal location encoding, defined as:

[0036] ;

[0037] ;

[0038] Where pos represents the time position, d represents the time embedding dimension, and i represents the dimension index of the position encoding;

[0039] Subsequently, the encoded temporal sequence is input into a multi-head self-attention layer. Through parallel computation of weighted aggregation of multiple attention heads, dependencies at different time scales are captured, defined as follows:

[0040] ;

[0041] ;

[0042] Where Q, K, and V represent the query matrix, key, and value, respectively; The dimension of the key is represented by Score, which represents the attention weight matrix. This represents the output of each I attention head;

[0043] Finally, the output of the multi-head self-attention layer is normalized by LayerNorm and fed into the feedforward network for feature extraction. The output of the feedforward network then yields the final temporal features. ;

[0044] II. Spatial Feature Extraction: The spatial feature extraction module is used to capture the spatial similarity of observation stations at different latitudes and longitudes. A spatial feature extraction method integrating geographic location information, spatial similarity, and graph convolutional neural networks is constructed. First, the input data obtains a preliminary spatial representation through a spatial feature stream, acquiring the latitude and longitude information of each station, i.e., spatial structure information. Then, geographic location is used to calculate... Determining the spatial relationships between stations involves calculating the spherical geographic distances between them. These spherical geographic distances are calculated using the Haversine formula, which is defined as follows:

[0045] ;

[0046] Where R represents the Earth's average radius; and These represent the latitude and longitude of the station, respectively.

[0047] Subsequently, a spatial weight matrix is ​​constructed based on the spherical geographic distance between stations, and a Gaussian kernel function is used to model the spatial similarity between stations, which is defined as:

[0048] ;

[0049] in, This represents a learnable scale parameter used to adjust the rate at which spatially dependent weights decay with distance; This represents the spherical geographic distance between station i and station j;

[0050] To combine the semantic features and spatial structure information of the monitoring stations, a learnable embedding encoding layer is designed to construct a learnable embedding vector for each monitoring station. Assume there are N monitoring stations in total, and each station... Associate an embedding vector Then the embedding matrix of all sites is represented as:

[0051] ;

[0052] Subsequently, combining the adjacency matrix and node features, spatial feature extraction is achieved through a graph convolutional neural network. The core feature update formula is as follows:

[0053] ;

[0054] in, This represents an adjacency matrix containing self-connections. Let it be its corresponding degree matrix. For the features of the l-th layer nodes, The weights of the graph convolutional layer, The activation function is used; the graph convolutional neural network effectively fuses information from the neighborhood of each site to model the spatial propagation dynamics and multi-scale correlations of sea surface temperature, in order to obtain the final spatial features. ;

[0055] III. Remote Sensing Feature Extraction: Considering the large-scale coverage of remote sensing data, a remote sensing feature extraction module is designed. First, the remote sensing data is fed into the remote sensing feature stream. Then, a linear feature projection network maps the original remote sensing observation data to a feature space compatible with ground observation features, forming an initial remote sensing feature representation.

[0056] ;

[0057] in, and These represent the trainable parameter matrix and the bias term, respectively.

[0058] A multi-source attention mechanism is introduced to enhance feature representation. Ground observation features are used as the query vector, and remote sensing features are used as the key K and value V. The attention calculation process is as follows:

[0059] ;

[0060] Here, d represents the dimension of the key vector; the multi-source attention mechanism captures highly correlated areas between ground and remote sensing information, dynamically enhances the remote sensing representation ability of important spatial locations, and captures the correlation between the two types of data through attention calculation; subsequently, the data after LayerNorm normalization is further extracted and enhanced through the feature enhancement layer, finally obtaining enhanced remote sensing features. ;

[0061] IV. In the conditional information construction stage, time characteristics are incorporated. Spatial features and remote sensing features Joint fusion is performed through a feature fusion layer to construct conditional information. Considering the varying degrees of dependence on different features across different time and space, an adaptive fusion module is designed, which uses a multilayer perceptron (MLP) to calculate the dynamic weights of each feature. Then, the data are weighted and fused to obtain the final conditional information, which is defined as:

[0062] ;

[0063] .

[0064] Optionally, the design of the loss function includes:

[0065] Design a loss function for multi-objective joint optimization, which consists of three key components: basic reconstruction loss, etc. Remote sensing data constraint loss and spatiotemporal consistency loss The basic reconstruction loss measures the difference between the model's predicted values ​​and the true values ​​at the missing locations, ensuring the model has basic data reconstruction capabilities. The remote sensing data constraint loss introduces remote sensing data as external supervision information to improve the model's inference ability in missing regions. The spatiotemporal consistency loss uses a smoothing loss function. Constraints are placed on the smoothness between predicted and known observations to enhance the continuity and consistency between interpolated and actual observation data; finally, the three losses are weighted by adjustable parameters. and By performing weighted combination, a joint loss function is constructed. :

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] Where N is the number of monitoring stations; This represents the interpolated value for the i-th station predicted by the model; This represents the actual value of the i-th station; This represents the target mask, which is the mask for the location to be predicted; This represents the remote sensing data value of the i-th station; This represents a conditional mask, used to assist in model training; Indicates the observation mask; and These represent the weights of remote sensing data constraint loss and spatiotemporal consistency loss, respectively.

[0071] In a second aspect, embodiments of the present invention provide a computer-readable storage medium comprising a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to execute the ocean spatiotemporal data interpolation method based on remote sensing conditional information diffusion in the first aspect or any possible implementation thereof.

[0072] Thirdly, embodiments of the present invention provide an electronic device, including: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions that, when executed by the device, cause the device to perform the ocean spatiotemporal data interpolation method based on remote sensing conditional information diffusion in the first aspect or any possible implementation of the first aspect.

[0073] The technical solution provided by this invention includes processing observation data from multiple ground stations and remote sensing data from L4 ocean to obtain observation datasets, remote sensing datasets, and datasets with different missing rates; standardizing and aligning the observation datasets and remote sensing datasets to obtain preprocessed data and a trained model; constructing an imputation architecture based on a conditional diffusion mechanism; and using the trained model to perform imputation inference for datasets with different missing rates. This method improves the accuracy of imputing missing values ​​in observation data and ensures the physical rationality and overall continuity of the imputation results. Attached Figure Description

[0074] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0075] Figure 1 A flowchart of a marine spatiotemporal data interpolation method based on remote sensing conditional information diffusion provided in an embodiment of the present invention;

[0076] Figure 2 A schematic diagram of an interpolation framework based on conditional diffusion mechanism provided in an embodiment of the present invention;

[0077] Figure 3 This is a schematic diagram of the interpolation architecture of the marine remote sensing conditional diffusion mechanism provided in an embodiment of the present invention;

[0078] Figure 4 A flowchart illustrating the construction of condition information provided in embodiments of the present invention;

[0079] Figure 5This is a schematic diagram of missing types of marine spatiotemporal data provided in an embodiment of the present invention, wherein (a) is missing random points, (b) is missing contiguous blocks, and (c) is missing mixed data.

[0080] Figure 6 The following is a comparison chart of results for different indicators provided in the embodiments of the present invention, wherein (a) is a comparison of results under the root mean square error indicator, (b) is a comparison of results under the mean absolute error indicator, and (c) is a comparison of results under the continuous graded probability score indicator.

[0081] Figure 7 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0082] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0083] It should be understood that the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0084] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” used in the embodiments of this invention are also intended to include the plural forms unless the context clearly indicates otherwise.

[0085] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0086] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0087] Figure 1 A flowchart of the ocean spatiotemporal data interpolation method based on remote sensing conditional information diffusion provided in the embodiments of the present invention is shown below. Figure 1 As shown, this invention is the first to apply the conditional diffusion mechanism to intelligent interpolation of marine spatiotemporal data, innovatively solving the challenges faced by traditional methods in processing marine data. The overall interpolation process of this invention includes three main stages: data acquisition and preprocessing, model training, and interpolation inference. The method includes:

[0088] Step 1: Process the observation data from multiple ground stations and the remote sensing data from the L4 ocean to obtain the observation dataset, the remote sensing dataset, and datasets with different missing rates.

[0089] In this embodiment of the application, the present invention first downloads observation data from multiple ground stations in the Mediterranean region and L4 ocean remote sensing data from the official website of the Copernicus Marine Environment Monitoring Service (CMEMS), and processes and aligns the two types of data respectively.

[0090] In this embodiment of the application, step 1 includes:

[0091] a. Observation data processing:

[0092] After acquiring the observation data, it was stored in the form of .nc files for multiple stations. First, it was read and converted to .csv format. Since the time field provided by CMEMS is the time difference from a specific time to the current time, it was uniformly converted into a standardized date and time representation. Second, the observation data consists of multiple depths and variables, and only the temperature TEMP variable at a depth of 0.5 m for each station was retained. Finally, missing values ​​were processed and set to NaN. Then, the data from all stations were merged to generate the observation data file, i.e., the observation dataset .csv. At the same time, the latitude and longitude information of each station was extracted to generate the latitude and longitude file, i.e., the latitude and longitude .csv. To improve the model's adaptability to different missing rates, missing datasets with different missing rates were manually constructed based on the real observation data.

[0093] b. Remote sensing data processing:

[0094] The remote sensing data for L4 ocean is hourly remote sensing grid data; based on the latitude and longitude information of each station, the data within the corresponding grid is extracted as the station's remote sensing data; since both the observation data and the remote sensing data are hourly data, they need to be aligned in the time dimension; finally, a remote sensing dataset, i.e., remote sensing dataset.csv, is generated.

[0095] Through the above processing, we obtain observation datasets, remote sensing datasets, and datasets with different missing rates, providing unified and standardized input data for subsequent model training and imputation inference.

[0096] In this embodiment of the application, hourly remote sensing grid data is scarce and difficult to obtain. Existing interpolation and reconstruction methods are mostly day-level. Therefore, this invention realizes fine-grained hourly data interpolation.

[0097] Step 2: Standardize and align the observation dataset and remote sensing dataset respectively to obtain preprocessed data and the trained model.

[0098] In this embodiment, the present invention first performs standardization and distribution alignment processing on the observation data and remote sensing data respectively, so as to reduce the differences in data sources and improve the training effect.

[0099] In this embodiment of the application, the preprocessed data obtained in step 2 includes:

[0100] Observational data standardization: Calculate the mean and standard deviation of the training set and normalize the original data using a standardization formula; Remote sensing data distribution alignment: Map the data to a distribution range similar to the observation data through statistical feature calculations to eliminate scale differences between different data sources.

[0101] When constructing training samples, a time step sequence is generated, and observation data masks and conditional masks are constructed. The observation data, remote sensing data, and corresponding masks are integrated to form the data batch required for training.

[0102] In this embodiment of the application, step 2 obtains the trained model, and its training process includes:

[0103] The preprocessed data is input into the interpolation model, forward propagation is performed to obtain the prediction results, and the loss between the predicted value and the true value is calculated. The loss function comprehensively considers the data reconstruction accuracy, remote sensing data constraints, and spatiotemporal consistency. The gradient is calculated through the backpropagation algorithm, and the model parameters are updated using the Adam optimizer. The above training process is iterated until the model converges. Finally, the trained model is saved.

[0104] Step 3: Construct an imputation architecture based on the conditional diffusion mechanism, and use the trained model to perform imputation inference on datasets with different missing rates.

[0105] In this embodiment, step 3 includes: during the inference phase, after loading the model, the input observation dataset and remote sensing dataset are first standardized and aligned in the same way as during the training phase to generate time series codes and corresponding masks; subsequently, an interpolation framework based on a conditional diffusion mechanism is constructed, such as... Figure 2As shown, its conditional diffusion mechanism includes:

[0106] h. During the forward diffusion phase, the model gradually adds Gaussian noise to the original sea surface temperature data, causing it to evolve into a form close to a standard normal distribution. This process constitutes a Markov chain, defined as:

[0107] ;

[0108] ;

[0109] in, Indicates diffusion time step Data; Indicates diffusion time step Data; This represents the preset noise scheduling parameters used to control the intensity of noise injection at each step; I represents the identity matrix. Represents the original data; This represents the data after adding noise at all time steps from 1 to T, i.e., the final noisy data; as the diffusion mechanism progresses, the original data structure is gradually lost, and the synchronously generated time series code and corresponding mask will be used for conditional control in the reverse diffusion stage;

[0110] i. In the inverse diffusion stage, the model starts from a pure Gaussian noise distribution and uses the inverse diffusion algorithm to iteratively denoise the noise; in each step, the model relies on conditional information. ,in Indicates time characteristics, Indicates spatial characteristics, Representing remote sensing features; the target data of the missing area is reconstructed step by step through a denoising process. The inverse diffusion process is defined as:

[0111] ;

[0112] in, This indicates that the noise prediction model is based on the data from the diffusion time step t. Time step t and condition information The predicted mean term; This represents the predefined learnable noise variance;

[0113] In each step, the model utilizes prior features to guide the current state, effectively reducing the uncertainty of data recovery. By integrating multi-source conditional information within this diffusion-reverse diffusion framework, the model can not only progressively restore missing sea surface temperature data, but also maintain consistency with known observations, spatial physical structures, and remote sensing priors during the recovery process, thereby improving the overall reconstruction quality.

[0114] j. In the post-processing stage, after completing the inverse diffusion process, the system de-standardizes the prediction results to map them back to the original data distribution range. Then, the system performs spatiotemporal consistency checks and physical rationality verification on the interpolation results to ensure that the prediction data and the original observation data are spatially and temporally consistent. Finally, the system integrates the interpolation results with the original observation data to generate complete and continuous spatiotemporal sequence data for data completion.

[0115] This invention effectively captures the spatiotemporal correlation of data through an iterative refinement mechanism of diffusion and dediffusion; simultaneously, it fully utilizes the complementary information of observational and remote sensing data, combined with feature fusion and multiple constraints, to achieve accurate prediction and reliable reconstruction of missing data. This method not only effectively solves the problem of missing marine data, but also has significant practical application value and innovative significance while ensuring the accuracy and reliability of the interpolation results.

[0116] In embodiments of the present invention, such as Figure 3 As shown, the interpolation architecture of the marine remote sensing conditional diffusion mechanism consists of an input layer, a conditional information construction module, a noise prediction module, and an output layer. This architecture is the first to apply the conditional diffusion mechanism based on remote sensing data to the field of marine spatiotemporal data interpolation. Through a specially designed marine approach, it can accurately model and interpolate marine spatiotemporal data.

[0117] First, in the forward diffusion phase, the model gradually injects noise into the original data, transforming it into noise distribution data. In the reverse diffusion phase, guided by conditional information, the model gradually reconstructs the target data through an iterative denoising process. To model the spatiotemporal characteristics of the ocean, a conditional information construction module is designed for triple feature extraction. Temporal and spatial features are extracted using Transformer and Graph Convolutional Neural Network (GNN), respectively. Remote sensing data is then combined to enhance observational information and extract remote sensing features. This design not only effectively extracts temporal dependencies and spatial correlations from ocean data but also effectively extracts remote sensing information to provide further data enhancement for ground observation data. The noise prediction model internally includes temporal attention, spatial attention, and a graph neural network (GNN). Multiple loss functions are used for optimization to ensure the accuracy of the generated interpolated data. Finally, the generated interpolated data is combined with the original observation data to form a complete, spatiotemporally continuous ocean dataset, laying the data foundation for subsequent ocean data processing and prediction.

[0118] In this embodiment of the invention, the conditional information includes triple feature extraction. The extracted triple features are dynamically fused to obtain the conditional information, which serves as an aid to improve the interpolation effect in the denoising process. Remote sensing data is used as the conditional information to reconstruct missing marine observation data, thereby generating a complete and continuous marine spatiotemporal observation dataset.

[0119] In embodiments of the present invention, such as Figure 4 As shown, the triple feature extraction and conditional information construction includes:

[0120] To comprehensively capture the spatiotemporal dynamics of marine environmental data, this study combines temporal feature extraction, spatial feature extraction, and feature encoding based on remote sensing feature extraction. It also incorporates conditional information to guide the inverse diffusion process in the diffusion model. Triple feature extraction is performed through these modules: temporal feature extraction captures the evolution of marine data across different time scales; spatial feature extraction characterizes the spatial correlation and similarity between observation stations; and remote sensing feature extraction incorporates large-scale remote sensing information to provide spatial background constraints for local observations. The synergistic effect of these three modules enables the model to comprehensively model across time, space, and multi-source information, conforming to the nonlinear distribution characteristics of marine spatiotemporal data and effectively improving the accuracy and stability of missing value imputation.

[0121] I. Temporal Feature Extraction: For the multi-scale temporal features of sea surface temperature series data, a temporal feature extraction module based on an improved Transformer structure is employed. First, the input data is fed into a temporal feature stream, and time points in the time series are mapped to a high-dimensional feature space through temporal location encoding, defined as:

[0122] ;

[0123] ;

[0124] Where pos represents the time position, d represents the time embedding dimension, and i represents the dimension index of the position encoding;

[0125] Subsequently, the encoded temporal sequence is input into a multi-head self-attention layer. Through parallel computation of weighted aggregation of multiple attention heads, dependencies at different time scales are captured, defined as follows:

[0126] ;

[0127] ;

[0128] Where Q, K, and V represent the query matrix, key, and value, respectively; The dimension of the key is represented by Score, which represents the attention weight matrix. This represents the output of each I attention head;

[0129] Finally, the output of the multi-head self-attention layer is normalized using LayerNorm and then fed into a feedforward network for feature extraction. The feedforward network consists of two linear transformations and activation functions, capable of independently processing each time step and extracting higher-level temporal features. The output of the feedforward network yields the final temporal features. ;

[0130] II. Spatial Feature Extraction: The spatial feature extraction module is used to capture the spatial similarity of observation stations at different latitudes and longitudes. A spatial feature extraction method integrating geographic location information, spatial similarity, and graph convolutional neural networks is constructed. First, the input data obtains a preliminary spatial representation through a spatial feature stream, acquiring the latitude and longitude information of each station, i.e., spatial structure information. Then, geographic location is used to calculate... Determining the spatial relationships between stations involves calculating the spherical geographic distances between them. These spherical geographic distances are calculated using the Haversine formula, which is defined as follows:

[0131] ;

[0132] Where R represents the Earth's average radius; and These represent the latitude and longitude of the station, respectively.

[0133] The above formula can accurately calculate the spherical positional relationship of each observation station on the Earth's surface, which can help in the subsequent construction of spatial map relationships.

[0134] Subsequently, a spatial weight matrix is ​​constructed based on the spherical geographic distance between stations, and a Gaussian kernel function is used to model the spatial similarity between stations, which is defined as:

[0135] ;

[0136] in, This represents a learnable scale parameter used to adjust the rate at which spatially dependent weights decay with distance; This represents the spherical geographic distance between station i and station j;

[0137] The dynamic weighting mechanism described above in this invention enables the model to adapt to the spatial propagation characteristics of different regions and to adapt to changes in spatial structure at different scales.

[0138] To combine the semantic features and spatial structure information of the monitoring stations, a learnable embedding encoding layer is designed to construct a learnable embedding vector for each monitoring station. Assume there are N monitoring stations in total, and each station... Associate an embedding vector Then the embedding matrix of all sites is represented as:

[0139] ;

[0140] The aforementioned embedding mechanism enables the model to better understand the geographical relationships between sites and capture potential differences in their functions and attributes.

[0141] The model's input includes geographic location information, namely the latitude and longitude data of each station; the adjacency matrix is ​​constructed from the latitude and longitude information of each station, representing spatial structure information; and the embedding matrix is ​​the embedding vector automatically learned by the model during training based on the adjacency matrix. Each station corresponds to an embedding matrix, representing semantic features.

[0142] Subsequently, combining the adjacency matrix and node features, spatial feature extraction is achieved through a graph convolutional neural network. The core feature update formula is as follows:

[0143] ;

[0144] in, This represents an adjacency matrix containing self-connections. Let it be its corresponding degree matrix. For the features of the l-th layer nodes, The weights of the graph convolutional layer, The activation function is used; the graph convolutional neural network effectively fuses information from the neighborhood of each site to model the spatial propagation dynamics and multi-scale correlations of sea surface temperature, in order to obtain the final spatial features. This provides physically consistent spatial prior information for feature fusion and generation tasks;

[0145] III. Remote Sensing Feature Extraction: Considering the large-scale coverage of remote sensing data, a remote sensing feature extraction module is designed. First, the remote sensing data is fed into the remote sensing feature stream. Then, a linear feature projection network maps the original remote sensing observation data to a feature space compatible with ground observation features, forming an initial remote sensing feature representation.

[0146] ;

[0147] in, and These represent the trainable parameter matrix and the bias term, respectively; this mapping can effectively alleviate the scale differences and complementarity between remote sensing data and ground data.

[0148] A multi-source attention mechanism is introduced to enhance feature representation. Ground observation features are used as the query vector, and remote sensing features are used as the key K and value V. The attention calculation process is as follows:

[0149] ;

[0150] Here, d represents the dimension of the key vector; the multi-source attention mechanism captures highly correlated areas between ground and remote sensing information, dynamically enhances the remote sensing representation ability of important spatial locations, and captures the correlation between the two types of data through attention calculation; subsequently, the data after LayerNorm normalization is further extracted and enhanced through the feature enhancement layer, finally obtaining enhanced remote sensing features. This mechanism can capture areas of high correlation between ground and remote sensing information, dynamically enhancing the remote sensing representation capabilities of important spatial locations.

[0151] IV. In the conditional information construction stage, time characteristics are incorporated. Spatial features and remote sensing features Joint fusion is performed through a feature fusion layer to construct conditional information. Considering the varying degrees of dependence on different features across different time and space, an adaptive fusion module is designed, which uses a multilayer perceptron (MLP) to calculate the dynamic weights of each feature. Then, the data are weighted and fused to obtain the final conditional information, which is defined as:

[0152] ;

[0153] .

[0154] In this embodiment of the invention, the adaptive fusion module dynamically weights and fuses the three extracted features, which is the core of conditional information construction. It extracts three features respectively to dynamically fuse and construct conditional information, and the conditional information assists in the denoising process for data interpolation.

[0155] In this embodiment of the invention, the design of the loss function includes:

[0156] Design a loss function for multi-objective joint optimization, which consists of three key components: basic reconstruction loss, etc. Remote sensing data constraint loss and spatiotemporal consistency loss The basic reconstruction loss measures the difference between the model's predicted values ​​and the true values ​​at the missing locations, ensuring the model has basic data reconstruction capabilities. The remote sensing data constraint loss introduces remote sensing data as external supervision information to improve the model's inference ability in missing regions. The spatiotemporal consistency loss uses a smoothing loss function. Constraints are placed on the smoothness between predicted and known observations to enhance the continuity and consistency between interpolated and actual observation data; finally, the three losses are weighted by adjustable parameters. and By performing weighted combination, a joint loss function is constructed. :

[0157] ;

[0158] ;

[0159] ;

[0160] ;

[0161] Where N is the number of monitoring stations; This represents the interpolated value for the i-th station predicted by the model; This represents the actual value of the i-th station; This represents the target mask, which is the mask for the location to be predicted; This represents the remote sensing data value of the i-th station; This represents a conditional mask, used to assist model training. It can be set manually to enable the model to obtain the location of known data. This represents the observation mask, used to distinguish whether original observation data exists, and it remains fixed. and These represent the weights of remote sensing data constraint loss and spatiotemporal consistency loss, respectively.

[0162] In this embodiment of the invention, the design of the multi-objective loss function not only ensures the prediction accuracy of the model in the target area, but also fully integrates the prior information of remote sensing data, and improves the smoothness and physical rationality of the interpolation results in the spatial and temporal dimensions through consistency constraints.

[0163] In this embodiment of the invention, the data sources and missing data types are as follows: The real data used comes from the official website of the Copernicus Marine Environment Monitoring Service. Observational data from Mediterranean surface buoy stations and their corresponding hourly remote sensing L4 data for the Mediterranean Sea are selected, covering hourly ocean surface temperature data for the entire year of 2023.

[0164] To more comprehensively evaluate the adaptability and interpolation effect of the method of this invention, three types of missing data were classified for the missing data in marine spatiotemporal data, such as... Figure 5 As shown in (a), (b), and (c), these represent random point missing, continuous block missing, and mixed missing, respectively.

[0165] Random point missing refers to a single missing point at a random location in the data, usually caused by equipment failure, communication interruption, or environmental factors. Point missing is quite common in practical applications, usually presenting as localized missing data, representing sporadic and irregular missing phenomena in marine observation data.

[0166] Contiguous block missing data refers to consecutive missing data over a period of time or within a spatial range, typically caused by prolonged equipment failure or environmental interference. Contiguous block missing data is usually concentrated in both time and space, exhibiting a relatively regular missing pattern, reflecting large-scale data loss due to systemic problems.

[0167] Mixed missing data combines the characteristics of point missing data and contiguous missing data, including both randomly distributed missing points and contiguous missing regions. Mixed missing data is more complex and typically occurs in actual ocean monitoring processes, representing data loss due to the simultaneous influence of multiple factors.

[0168] By dividing ocean spatiotemporal data into three categories, we can more comprehensively consider different patterns under missing ocean spatiotemporal data, and ensure that the imputation effect can meet actual needs when dealing with different missing situations.

[0169] In this embodiment of the invention, data partitioning and experimentation are conducted as follows: To avoid information leakage and seasonal bias, this experiment selects several representative months for training (e.g., January, February, April, May, July, August, October, and November), and divides February, May, August, and November into the last few months according to their chronological order. As a validation set, the rest The training set is reserved for the test set, which consists of months mutually exclusive with training and validation (e.g., March, June, September, and December). To avoid leakage of missing patterns, the historical window used to generate the condition mask during training is drawn only from months mutually exclusive with evaluation. This partitioning covers both intra-year seasonality and cross-monthly distribution differences, enabling stable evaluation of the model's generalization ability and robustness while ensuring strict temporal causality.

[0170] For the three types of missing data mentioned above, the model of this invention demonstrates greater accuracy compared to pure temporal imputation models and pure spatiotemporal imputation models at different missing rates. This is because the model of this invention not only considers the temporal and spatial factors of the ocean but also uses remote sensing data as augmentation and selects a diffusion model to avoid error accumulation. Furthermore, as... Figure 6 As shown in (a), (b), and (c), the horizontal axis represents different missing types corresponding to different missing rates, and the vertical axis represents the corresponding indicators (Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Continuous Ranking Probability Score (CRPS). The results of models with and without remote sensing data augmentation are compared and visualized under different indicators. It can be seen that using remote sensing data as augmentation can effectively improve imputation accuracy, especially when 50% of continuous blocks are missing, the indicator reduction exceeds [percentage missing]. This demonstrates the accuracy and practicality of the model in this invention.

[0171] This invention addresses the severe deficiencies and incompleteness in marine observation data. It employs a diffusion-reverse diffusion generation model to interpolate missing values ​​in observation data, using remote sensing data as conditional information. Furthermore, it constructs an intelligent reconstruction framework for multi-site spatiotemporal data interpolation tasks in the ocean, aiming to solve the spatial dependence problem of existing interpolation methods; the error accumulation problem of existing interpolation models; and the lack of an effective mechanism for utilizing the conditional information of remote sensing data in hourly high-frequency interpolation methods. First, this invention extracts the time-series features from in-situ site observation data, while simultaneously introducing remote sensing data as conditional information to supplement spatial distribution constraints. Remote sensing data provides a large-scale spatial background for the model, while site data provides high-resolution temporal evolution information. Second, it constructs an interpolation framework based on a conditional diffusion model. Through a diffusion process, missing data is perturbed into a noise distribution, and then, through a reverse diffusion process guided by remote sensing conditional information, missing values ​​are gradually restored, thereby achieving fitting of complex spatiotemporal distribution patterns. Compared with traditional methods, this method not only characterizes temporal dependence but also fully models the spatial correlation between sites and effectively suppresses the error accumulation problem caused by high missing rates.

[0172] The method of this invention makes full use of the spatial breadth of remote sensing data and the temporal depth of observation data. It uses hourly remote sensing data as input for conditional information in the interpolation process, and combines the extraction of temporal and spatial features of observation data to learn the spatiotemporal distribution pattern of ocean data. Through two processes of diffusion and de-diffusion, it achieves the interpolation of missing data, thereby forming a continuous and complete ocean observation dataset. This provides a solid foundation for subsequent analysis and modeling, and significantly improves data utilization efficiency and processing accuracy.

[0173] The interpolation method of this invention significantly improves interpolation accuracy and stability, fully utilizing remote sensing data as conditional information, and exhibits superior interpolation performance under different missing rates and missing modes. Especially under a high missing rate of 50%, it significantly improves the accuracy and stability of data recovery. This invention breaks through the bottleneck of high-resolution data interpolation, introducing remote sensing L4 hourly data as conditional information into the diffusion model for the first time. This enables the model to better respond to short-period phenomena in the marine environment, such as sudden changes, anomalies, and extreme weather, improving the usability of marine monitoring data in high-frequency applications such as real-time monitoring and short-term forecasting. This invention enhances cross-regional adaptability and generalization by designing a triple feature fusion framework for marine spatiotemporal data, combining the periodic changes in the time dimension, the geographical adjacency relationships in the spatial dimension, and the global spatial coverage capability of remote sensing data. This method models the spatial correlation between stations through a graph neural network (GNN), without relying on prior knowledge of specific regions, and can adapt to the spatial distribution of different regions. It can be extended to different sea areas and multivariate tasks, and has global application potential.

[0174] Compared with the prior art, the present invention has the following advantages:

[0175] (1) This invention applies the conditional diffusion model based on remote sensing data to the task of interpolating marine multi-buoy spatiotemporal data. Combined with a specially designed marine diffusion architecture, it uses the diffusion-reverse diffusion mechanism to achieve high-quality data recovery and uses remote sensing data as conditional information to enhance the modeling of multi-site observation data. This effectively solves the error accumulation problem of existing models and realizes accurate modeling and interpolation of marine spatiotemporal data.

[0176] (2) By introducing hourly L4 remote sensing data as an auxiliary condition, this invention overcomes the technical bottleneck of traditional interpolation based on daily-scale remote sensing data, supporting fine-grained interpolation of ocean data on an hourly basis. Compared with previous methods that mainly relied on daily or coarser-grained remote sensing data for missing value repair, this invention can handle higher-frequency hourly data, significantly improving temporal resolution. Under the fusion framework of in-situ observation data as the main body and remote sensing data as an auxiliary, the rationality, accuracy, and stability of interpolation are effectively improved. Compared with existing interpolation methods that solely rely on in-situ observations, this invention demonstrates stronger stability and robustness, thereby significantly improving the utilization efficiency and reliability of ocean observation data.

[0177] (3) This invention proposes a triple feature extraction and fusion mechanism: the temporal feature extraction module characterizes the temporal evolution, the spatial feature extraction module models the station correlation, and the remote sensing feature extraction module provides large-scale background constraints. The collaboration of the three not only makes up for the shortcomings of existing methods in spatial modeling, but also significantly improves the accuracy and robustness of interpolation.

[0178] (4) A multi-objective joint loss function suitable for marine data interpolation tasks was designed, including basic reconstruction loss, remote sensing data constraint loss and spatiotemporal consistency loss, to ensure the physical rationality and overall continuity of the interpolation results and significantly improve the performance of the model in marine data interpolation tasks.

[0179] The technical solution provided by this invention includes processing observation data from multiple ground stations and remote sensing data from L4 ocean to obtain observation datasets, remote sensing datasets, and datasets with different missing rates; standardizing and aligning the observation datasets and remote sensing datasets to obtain preprocessed data and a trained model; constructing an imputation architecture based on a conditional diffusion mechanism; and using the trained model to perform imputation inference for datasets with different missing rates. This method improves the accuracy of imputing missing values ​​in observation data and ensures the physical rationality and overall continuity of the imputation results.

[0180] The various steps in the embodiments of the present invention can be performed by an electronic device. This electronic device includes, but is not limited to, tablet computers, portable PCs, and desktop computers.

[0181] This invention provides a computer-readable storage medium including a stored program, wherein, when the program is running, it controls the electronic device containing the computer-readable storage medium to execute the above-described embodiment of the marine spatiotemporal data interpolation method based on remote sensing conditional information diffusion.

[0182] Figure 7 A schematic diagram of an electronic device provided in an embodiment of the present invention, such as... Figure 7 As shown, the electronic device 21 includes a processor 211, a memory 212, and a computer program 213 stored in the memory 212 and executable on the processor 211. When the computer program 213 is executed by the processor 211, it implements the marine spatiotemporal data interpolation method based on remote sensing conditional information diffusion in the embodiment. To avoid repetition, it will not be described in detail here.

[0183] Electronic device 21 includes, but is not limited to, processor 211 and memory 212. Those skilled in the art will understand that... Figure 7 This is merely an example of electronic device 21 and does not constitute a limitation on electronic device 21. It may include more or fewer components than shown, or combine certain components, or different components. For example, electronic device may also include input / output devices, network access devices, buses, etc.

[0184] The processor 211 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0185] The memory 212 can be an internal storage unit of the electronic device 21, such as a hard disk or RAM of the electronic device 21. The memory 212 can also be an external storage device of the electronic device 21, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, or FlashCard equipped on the electronic device 21. Furthermore, the memory 212 can include both internal and external storage units of the electronic device 21. The memory 212 is used to store computer programs and other programs and data required by network devices. The memory 212 can also be used to temporarily store data that has been output or will be output.

[0186] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0187] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for spatiotemporal data interpolation of the ocean based on remote sensing condition information diffusion, characterized in that, The method comprises: Step 1, processing the observation data of multiple ground sites and the remote sensing data of L4 ocean, to obtain an observation data set, a remote sensing data set, and data sets with different missing rates; Step 2, standardizing and aligning the distribution of the observation data set and the remote sensing data set respectively, to obtain preprocessed data, and to obtain a trained model; Step 3, constructing an interpolation architecture based on a conditional diffusion mechanism, and using the trained model to perform interpolation reasoning on data sets with different missing rates; The trained model obtained in step 2 has a training process comprising: The preprocessed data is input into the interpolation model for forward propagation to obtain a prediction result, and the loss between the predicted value and the true value is calculated; the loss function comprehensively considers the data reconstruction accuracy, the remote sensing data constraint, and the spatiotemporal consistency; the gradient is calculated by the back propagation algorithm, and the model parameters are updated by the Adam optimizer; the above training process is iterated until the model converges, and finally the trained model is saved; The interpolation architecture of the marine remote sensing conditional diffusion mechanism comprises an input layer, a conditional information construction module, a noise prediction module, and an output layer; Firstly, in the forward diffusion stage, the model gradually injects noise into the original data to convert it into noise distribution data; in the reverse diffusion stage, the model gradually reconstructs the target data through an iterative denoising process under the guidance of the conditional information; to model the spatiotemporal characteristics of the ocean, a conditional information construction module is designed to extract three features, i.e., time feature, space feature, and remote sensing feature, through a Transformer and a graph convolutional neural network, respectively; the noise prediction model includes time attention, space attention, and a graph neural network (GNN) inside; multiple loss functions are used for optimization, and finally the generated interpolation data is combined with the original observation data to form a complete and spatiotemporally continuous ocean data set.

2. The method of claim 1, wherein, The step 1 comprises: a. Observation data processing: After obtaining the observation data, store it in the form of multiple site.nc files; first, read and convert it to.csv format; unify it into a standardized date-time representation; second, the observation data consists of multiple depths and variables, only keep the temperature (TEMP) variable at 0.5 m depth of each site; finally, handle the missing values by setting them to nan; then merge the data of all sites to generate an observation data file, i.e., observation data set.csv; at the same time, extract the latitude and longitude information of each site to generate a latitude and longitude file, i.e., latitude.csv; to improve the model's adaptability to different missing rates, artificially construct missing data sets with different missing rates based on the real observation data; b. Remote sensing data processing: The remote sensing data of L4 ocean is hourly remote sensing grid data; according to the latitude and longitude information of each site, extract the data within the corresponding grid as the remote sensing data of the site; since both the observation data and the remote sensing data are hourly data, they need to be aligned in the time dimension; finally, generate a remote sensing data set, i.e., remote sensing data set.csv; Through the above processing, the observation data set, the remote sensing data set, and the data sets with different missing rates are obtained.

3. The method of claim 2, wherein, The pre-processed data obtained in step 2 includes: Observation data standardization: calculate the mean and standard deviation of the training set, and normalize the original data using the standardization formula; remote sensing data distribution alignment: through statistical feature calculation, it is mapped to a similar distribution range as the observation data to eliminate the scale difference between different data sources; When constructing the training sample, generate the time step sequence, and construct the observation data mask and the condition mask, integrate the observation data, remote sensing data and the corresponding mask to form the data batch required for training.

4. The method of claim 1, wherein, Step 3 includes: in the inference stage, first load the model and then perform standardization and distribution alignment processing on the input observation data set and remote sensing data set consistent with the training stage, generate time series encoding and corresponding mask; then, build an interpolation framework based on the conditional diffusion mechanism, which includes: h、In the forward diffusion stage, the model gradually evolves into a form close to the standard normal distribution by gradually adding Gaussian noise to the original sea surface temperature data, and this process constitutes a Markov chain, defined as: ； ； wherein, denotes the data at diffusion time step ; denotes the data at diffusion time step ; denotes a preset noise schedule parameter for controlling the intensity of noise injection at each step; I denotes an identity matrix; denotes the original data; denotes the data after adding noise from all time steps from 1 to T, i.e., the final noise data; as the diffusion mechanism progresses, the original data structure is gradually lost, and the time series encoding and the corresponding mask generated synchronously will be used for conditional control in the inverse diffusion stage. i. In the inverse diffusion stage, the model starts from a pure Gaussian noise distribution and iteratively denoises step by step using the inverse diffusion algorithm; in each step, the model relies on conditional information wherein denotes the temporal feature, denotes the spatial feature, denotes the remote sensing feature; the target data of the missing area is gradually reconstructed through a step-by-step denoising process, and the inverse diffusion process is defined as: ； wherein, denotes the data at diffusion time step t predicted by the noise prediction model , the time step t and the conditional information the mean term predicted; denotes a predefined learnable noise variance; j、In the post-processing stage, after completing the inverse diffusion process, the system performs inverse standardization on the prediction results to map them back to the original data distribution range; then, the interpolation results are checked for spatio-temporal consistency and physically verified to ensure that the predicted data is consistent with the original observation data in space and time; finally, the system integrates the interpolation results with the original observation data to generate complete and continuous spatio-temporal sequence data for data completion.

5. The method of claim 1, wherein, Triple feature extraction and condition information construction includes: To fully capture the spatio-temporal dynamic characteristics of marine environmental data, combine time feature extraction, spatial feature extraction, and feature encoding based on remote sensing feature extraction, and construct condition information to guide the inverse diffusion process in the diffusion model; triple feature extraction is performed through the time feature extraction module, spatial feature extraction module, and remote sensing feature extraction module. The time feature extraction module is used to capture the evolution law of marine data at different time scales; the spatial feature extraction module is used to describe the spatial correlation and similarity between observation stations; the remote sensing feature extraction module introduces large-scale remote sensing information to provide spatial background constraints for local observations; Ⅰ、Time feature extraction: for the multi-scale time series features of sea surface temperature sequence data, an improved Transformer structure-based time feature extraction module is used; first, input the data into the time feature flow, and map the time points in the time series to a high-dimensional feature space through time position encoding, which is defined as: ； ； Where pos represents the time position, d represents the time embedding dimension, and i represents the dimension index of the position encoding; Then, input the encoded time series into the multi-head self-attention layer to capture different time scale dependencies through parallel calculation of multiple attention heads, which is defined as: ； ； Wherein Q, K and V represent query matrix, key and value respectively; represents the dimension size of the key, and Score represents the attention weight matrix. represents the output of each I attention head; Finally, the output of the multi-head self-attention layer is standardized by layer normalization, and is sent to a feedforward network for feature extraction. The output of the feedforward network is the final temporal feature. ; II. Spatial feature extraction: The spatial feature extraction module is used to capture the spatial similarity of different latitude and longitude observation stations; the spatial feature extraction of the fusion of geographic location information, spatial similarity and graph convolutional neural network is constructed. First, the input data obtains the preliminary spatial representation through the spatial feature flow, and obtains the latitude and longitude information of each station, i.e. the spatial structure information; then, the geographic location is calculated to determine the spatial relationship between stations, which involves calculating the spherical geographic distance between stations. The spherical geographic distance between stations is calculated by the Haversine formula, which is defined as: ； where R represents the mean radius of the earth; and respectively represent the latitude and longitude of the site. Then, construct a spatial weight matrix based on the spherical geographic distance between stations, and use a Gaussian kernel function to model the spatial similarity between stations, which is defined as: ； wherein, denotes a learnable scale parameter for adjusting the decay speed of the spatial dependent weight with distance; denotes the spherical geographic distance between i-site and j-site; To combine the semantic features and spatial structure information of the sites, a learnable embedding coding layer is designed to construct a learnable embedding vector for each detection site. Assuming that there are N monitoring sites, each site is associated with an embedding vector The embedding matrix of all sites is represented as: ； Subsequently, combined with the adjacency matrix and node features, the spatial feature extraction is realized through the graph convolutional neural network, and the core feature updating formula is: ； wherein, denotes a self-connected adjacency matrix, is its corresponding degree matrix, is the feature of the l-th layer node, is the weight of the graph convolution layer, is an activation function; the graph convolutional neural network effectively fuses the information of each site neighborhood, models the propagation dynamics and multi-scale correlation of sea surface temperature in space, to obtain the final spatial features ; III. Remote sensing feature extraction: In view of the large-scale coverage characteristics of remote sensing data, a remote sensing feature extraction module is designed. First, the remote sensing data is input into the remote sensing feature stream as input, and then the original remote sensing observation data is mapped to a feature space compatible with the ground observation features through a linear feature projection network to form an initial remote sensing feature representation: ； wherein, and respectively denote trainable parameter matrices and bias terms; A multi-source attention mechanism is introduced to enhance feature expression. The ground observation features are used as query vectors, and the remote sensing features are used as keys K and values V. The attention calculation process is: ； wherein d represents the dimension of the key vector; the multi-source attention mechanism captures the high correlation area between the ground and remote sensing information, dynamically enhances the remote sensing expression ability of important spatial positions, and captures the correlation between the two kinds of data through attention calculation; subsequently, the data passing through the layer normalization LayerNorm is further extracted and enhanced through the feature enhancement layer, and finally the enhanced remote sensing features are obtained ; IV. In the condition information construction stage, the time characteristics , spatial characteristics and remote sensing characteristics are jointly fused through a feature fusion layer to construct condition information ; considering that different time and space have different dependence on various characteristics, an adaptive fusion module is designed to calculate the dynamic weight of each feature through a multi-layer perception (MLP) , and then the final condition information is obtained by weighted fusion, which is defined as: ； 。 6. The method of claim 1, wherein, The design of the loss function includes: A multi-objective joint optimization loss function is designed, which is composed of three key components, namely, the basic reconstruction loss , the remote sensing data constraint loss , and the spatio-temporal consistency loss ; the basic reconstruction loss is used to measure the difference between the model prediction value and the true value at the missing position, to ensure that the model has basic data reconstruction capability; the remote sensing data constraint loss introduces remote sensing data as external supervision information to improve the inference ability of the model in the missing area; the spatio-temporal consistency loss uses a smoothing loss function to constrain the smoothness between the prediction value and the known observation value, to enhance the continuity and consistency between the interpolated data and the actual observation data; finally, the three losses are combined by adjustable weight parameters and to form a joint loss function : ； ； ； ； where N is the number of monitoring sites; represents the interpolated value of the i-th site predicted by the model; represents the true value of the i-th site; represents the target mask, which is the mask of the position that needs to be predicted; represents the remote sensing data value of the i-th site; represents the conditional mask, which is used to assist the model training; represents the observation mask; and respectively represent the weight of the remote sensing data constraint loss and the spatio-temporal consistency loss.

7. A computer-readable storage medium, characterized in that, The computer readable storage medium includes a stored program, wherein the program controls the device where the computer readable storage medium is located to execute the ocean spatio-temporal data interpolation method based on remote sensing conditional information diffusion according to any one of claims 1 to 6 when the program is running.

8. An electronic device, comprising: including: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs include instructions that, when executed by the device, cause the device to execute the ocean spatio-temporal data interpolation method based on remote sensing conditional information diffusion according to any one of claims 1 to 6.

Images