Forecasting method and system for observing dense space-time physical field based on atmosphere and ocean points

Abstract

The invention relates to the technical field of observation data processing, in particular to a forecasting method and system for observing a dense space-time physical field based on atmosphere and ocean points. Obtaining observation data, wherein the observation data comprises historical dense reanalysis space-time physical field data and corresponding sparse sensor observation data; performing data preprocessing based on the obtained observation data; carrying out vectorization representation learning according to the preprocessed data; performing intermediate feature extraction based on a representation learning result; and on the basis of the extracted intermediate features and embedding spatio-temporal information, a dense spatio-temporal physical field of the forecast is formed. According to the method, the uncertainty from sparse sensor observation data to dense space-time physical field mapping is greatly reduced, so that the forecasting process is more stable and robust.

Description

Technical Field

[0001] This invention relates to the field of observation data processing technology, and in particular to a prediction method and system based on dense spatiotemporal physical fields observed from atmospheric and oceanic points. Background Technology

[0002] In fields such as oceanography, meteorology, and climate science, accurately acquiring high-resolution geophysical fields (such as sea surface wind fields, wave fields, and temperature fields) is crucial for scientific research, shipping safety, and disaster early warning. While traditional numerical weather prediction (NWP) models can provide forecasts, they are typically computationally expensive, have long processing times, and are highly sensitive to the quality of the initial field. On the other hand, direct observation methods, such as satellite remote sensing, are limited by factors such as cloud cover; and while in-situ sensors (such as ocean sensors and land-based weather stations) can provide high-precision time-series data, their spatial coverage is extremely sparse.

[0003] Therefore, how to economically, efficiently, and accurately predict a complete and dense spatiotemporal physical field from limited and sparse sensor data has been a long-standing and challenging scientific and engineering problem. Mathematically, this problem is often considered an ill-posed problem because there are infinitely many possible solutions to the mapping from sparse input to dense output. Existing data-driven methods, such as simple interpolation or traditional machine learning regression models, often struggle to capture complex nonlinear spatiotemporal dynamics, leading to overly smooth and ambiguous predicted spatiotemporal physical fields, or poor performance under extreme conditions. Summary of the Invention

[0004] To address the problem that existing technologies struggle to capture complex nonlinear spatiotemporal dynamics when processing mappings from sparse inputs to dense outputs, resulting in overly smooth or ambiguous predicted spatiotemporal physical fields, or poor performance under extreme conditions, this invention provides a prediction method and system based on atmospheric and oceanic point observations of dense spatiotemporal physical fields.

[0005] In a first aspect, the present invention provides a prediction method based on dense spatiotemporal physical fields observed at atmospheric and oceanic points, employing the following technical solution: A prediction method based on dense spatiotemporal physical fields observed at atmospheric and oceanic points, comprising: Acquire observational data, including historical dense reanalysis spatiotemporal physical field data and corresponding sparse sensor observation data; Data preprocessing is performed based on the acquired observation data; Vector representation learning is performed based on the preprocessed data; Intermediate feature extraction is performed based on representation learning results; Based on the extracted intermediate features and embedded spatiotemporal information, a predicted dense spatiotemporal physical field is formed.

[0006] Secondly, a forecasting system based on dense spatiotemporal physical fields observed at atmospheric and oceanic points includes: The data acquisition module is configured to acquire observation data, including historical dense reanalysis spatiotemporal physical field data and corresponding sparse sensor observation data. The data preprocessing module is configured to perform data preprocessing based on the acquired observation data; The representation learning module is configured to perform vectorized representation learning based on the preprocessed data. The feature extraction module is configured to extract intermediate features based on the representation learning results. The sequence prediction module is configured to perform sequence modeling based on feature extraction data and representation learning results; The field formation module is configured to form a predicted dense spatiotemporal physical field based on extracted intermediate features and embedded spatiotemporal information.

[0007] Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for predicting dense spatiotemporal physical fields based on atmospheric and oceanic point observations.

[0008] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a prediction method for dense spatiotemporal physical fields based on atmospheric and oceanic point observations.

[0009] In summary, the present invention has the following beneficial technical effects: 1. Overcoming ill-posedness problem: By constraining the solution space within a finite, learned discrete potential representation space, this invention greatly reduces the uncertainty of mapping from sparse sensor observation data to dense spatiotemporal physical fields, making the prediction process more stable and robust.

[0010] 2. High efficiency and real-time performance: Once the model is trained, inference only involves lightweight sequence prediction and decoding processes, with computational costs far lower than traditional numerical simulation methods, enabling near real-time prediction and forecasting of spatiotemporal physical fields.

[0011] 3. Data efficiency and cost-effectiveness: This method can achieve high-quality dense spatiotemporal physical field forecasts using only a small amount of sparse sea surface meteorological sensor data (e.g., spatial coverage of less than 5%), which significantly reduces the dependence on expensive and dense observation networks.

[0012] 4. Broad Applicability: The core architecture of this invention is universal and not limited to specific physical fields or sensor types. By replacing the corresponding datasets, this method can be applied to multiple geophysical science fields such as oceanography, atmosphere, and geology. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of a prediction method based on dense spatiotemporal physical fields observed at atmospheric and oceanic points, according to Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the structure of Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the dynamic modal fusion encoder DMFE according to Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the lightweight hierarchical attention network LHAN according to Embodiment 1 of the present invention; Figure 5 This is the overall framework diagram of Embodiment 2 of the present invention; Figure 6 This is a wind field UV component forecast effect diagram of Embodiment 1 of the present invention; Figure 7 This is a wind speed forecast effect diagram of Embodiment 1 of the present invention. Detailed Implementation

[0014] The present invention will be further described in detail below with reference to the accompanying drawings.

[0015] Example 1 It should be emphasized that in this invention, the observation data is sparse and the forecast data is dense, but the intermediate process does not go through NWP, but uses artificial intelligence methods for rapid forecasting. The point observation and sparse observation mentioned in this application refer to the same type of concept.

[0016] Reference Figure 1 This embodiment of a prediction method based on dense spatiotemporal physical fields observed at atmospheric and oceanic points includes: S1. Acquire historical dense reanalysis spatiotemporal physics data and corresponding sparse sensor observation data, including: S11. Acquire a historical grid dataset containing reanalysis spatiotemporal physical field data within a predefined target geographic area, as well as actual time-series observation data containing corresponding meteorological variables from multiple spatially discrete sensor stations deployed within that area.

[0017] S12. The historical gridded dataset can be obtained from public climate data services or generated through numerical models, and the sparse sensor station observation data can be obtained from national or international marine data centers. Data selection should ensure a sufficiently long timeframe to cover a variety of typical environmental conditions, thereby guaranteeing adequate model training. Subsequently, depending on the specific circumstances, the samples are divided into training, validation, and test sets.

[0018] S2. Preprocess sensor observation data and construct time-series training samples, including: S21. First, data preprocessing is performed to improve data quality. Statistical methods (such as threshold judgment based on mean and standard deviation) or machine learning algorithms (such as the Isolation Forest algorithm) are used to filter out missing and outlier values ​​in sensor observations. At the same time, invalid data with incorrect formats is identified and corrected through program validation rules. The raw data is converted into a unified format that meets the needs of mathematical model analysis (such as standardized data structures to ensure data consistency), laying the foundation for subsequent analysis.

[0019] Specifically, outlier removal and correction are performed using the 3σ rule based on a sliding window. For the sensor's time-series data X, the sliding window length is set to L. For the center point within the window... Calculate the mean of the data within the window. and standard deviation If satisfied Then determine These are outliers. For data points identified as outliers or original missing data points, linear interpolation is used to fill them in. , in, These are the corrected values. and These are the adjacent valid observations.

[0020] S22. Based on the time alignment principle, through precise timestamp matching and interpolation (if there is a slight time deviation), sparse sensor data that matches the data is selected to ensure consistency and correspondence between the two in the time dimension, so as to effectively realize the nonlinear mapping between the two types of data. The data is standardized to complete the data preprocessing and divided into training set, validation set and test set.

[0021] The specific time alignment calculation process is as follows: Let the timestamp sequence of the reanalysis grid dataset be... The timestamp sequence of sparse sensor observation data is Set the maximum allowable time deviation threshold to . (e.g., 10 minutes). For any reanalysis data time point Find the moment with the smallest time difference in the sensor sequence. ,Right now .like Then directly use the corresponding sensor observation value Matching to time .like This indicates that sensor data is missing at that moment, and it needs to be utilized. The two most recent valid sensor times Perform weighted interpolation: , This allows for precise alignment of sparse and dense grid data across the time dimension.

[0022] S23. Construct training sample pairs, Let the time series be... T is the total number of time steps. At time t as the baseline, the corresponding sample window data W contains n time steps, i.e. , It is the dense reanalysis physical field prediction data at time t, forming an effective accumulation of historical information about the state at time t; S24. Z-Score normalization is a commonly used data preprocessing method. By standardizing the input and target features, it can eliminate differences in data units, which helps improve the training efficiency and convergence speed of the model. This method transforms the original data into a distribution with a mean of 0 and a standard deviation of 1, thereby making the scales of features comparable. However, this invention is not limited to any specific normalization method; any data scaling technique that can improve the stability of model training is applicable.

[0023] S241. In this invention, for each feature i, its mean on the training set is first calculated. and standard deviation Then for each data point The standardization process is performed, and the specific formula is as follows: , Where i represents the feature number, and h and w represent the data points in the latitude and longitude directions, respectively. The standardized data has a distribution with a mean of 0 and a standard deviation of 1.

[0024] S242. In the model prediction phase, in order to restore the standardized data to the original scale, the output results are destandardized, as shown in the following formula: , In this invention, the Z-Score normalization method is used for both input and target in the self-prediction training stage and the second training stage of the dense physics field prediction. This not only improves the stability of model training, but also effectively enhances the model's adaptability to changes in data scale, making the convergence speed of iterative algorithms such as gradient descent faster and reducing training time.

[0025] Reference Figure 2 , S3. A discrete latent representation space and a decoder are learned through a first-stage model (which can be understood as a vector quantization representation learning system), achieved through a representation learning module. In a preferred embodiment, this module may employ an improved vector quantization variational autoencoder architecture, including: Specifically, historical dense geophysical field data samples are input into the encoder portion of this module. In a preferred embodiment, the encoder may be implemented by a convolutional neural network, which functions to map the input gridded data into a continuous intermediate feature representation. However, the invention is not limited thereto; other neural network architectures capable of effectively extracting spatial features, such as graph neural networks (GNNs) or various autoencoders, may also be used to perform this function.

[0026] The continuous feature is then mapped to a discrete latent representation space through a quantization layer. This space consists of a finite set of learnable embedding vectors (basis vectors).

[0027] Finally, the decoder part of this module is responsible for predicting the quantized discrete representation as a dense physical field that is as consistent as possible with the original input. The goal of training is to minimize the prediction error and other regularization terms, thereby ensuring that the discrete space can effectively capture and represent the key patterns of the real spatiotemporal physical field. The specific implementation of the model algorithm is explained below: S31. First is the encoder module, where the input tensor shape is... The reanalysis of dense spatiotemporal physical field prediction data. Represents batch size. The number of time series representing continuous time resolution. This represents the number of multi-element channels, covering core meteorological variables such as sea surface wind field, wave field, and temperature field. Represents the height of the feature map. Represents the width of the feature map. and The value must match the spatial grid division standard of the target area.

[0028] To adapt to the input characteristics of subsequent convolutional neural networks and enhance the features of periods of drastic change such as severe convective weather, the tensor of the original observation data was first processed. Perform variance-based weighted reorganization. Time-series variance calculation: for each sample... At time step Data slices Calculate its spatial variance This characterizes the intensity of the physical field at that moment (such as the gradient change at the center of a typhoon): , Adaptive weight generation: Calculating reinforcement weights at each time step based on variance proportions And introduce basic weight coefficients (For example =0.5) Prevent information loss during low variance periods: , Weighted merging and dimensional reshaping: Applying weights to the data at the corresponding time step and merging the time dimension T with the channel dimension. Perform stacking and merging to generate a new model input tensor. : , The final result As input to subsequent networks, it effectively highlights the physical characteristics of high-dynamic periods.

[0029] In the multi-level feature extraction unit of the encoder, the specific calculation process is as follows: Assume the first... The input features of the layer are First, features from different receptive fields are captured using multi-scale convolutional layers. Let's assume... Groups of convolutional kernels of different sizes (e.g.) The feature maps are output separately and then concatenated and fused: , Subsequently, after normalization and activation: , Finally, the output is weighted by an attention sublayer (such as SE-Block or spatial attention): , in This represents element-wise multiplication. The function is a sigmoid function, and the last term represents the residual connection. After multiple stacked layers, the final output is a continuous feature representation. .

[0030] From a mathematical mapping perspective, this process can be represented as: , in, The mapping function representing the encoder is a composite function composed of multiple layers of feature extraction units. In actual operation, this function first extracts the basic texture features of the spatiotemporal physical field through shallow units, then captures mesoscale structural features through mid-level units, and finally aggregates global distribution features through deep units, gradually transforming the input data x into a high-dimensional feature vector. It fully preserves the local details and global distribution characteristics of the spatiotemporal physical field, providing high-quality feature input for the subsequent quantization process.

[0031] S32. Next is the vector quantization module. This module acts as a bridge connecting the encoder's continuous feature space and discrete symbol space. Its core objective is to represent the continuous features output by the encoder using a nearest neighbor search mechanism. This process maps features into a predefined discrete latent space. This achieves efficient discretization of features, filtering out redundant noise in the input data and reducing the computational complexity of subsequent prediction models.

[0032] The specific quantization calculation process is as follows: Suppose the discrete latent space consists of a set of learnable embedding vectors, denoted as... ,in The size of the discrete latent space (i.e., codebook capacity, for example) = 512). Each embedding vector Dimensions It must be strictly consistent with the dimension of the continuous feature vector output by the encoder to ensure algebraic compatibility of the mapping process.

[0033] For the feature map output by the encoder Each spatial location in eigenvectors Calculate its relationship with all embedding vectors in the codebook. The difference is considered. To balance the similarity of feature values ​​(Euclidean distance) and the consistency of feature orientation (cosine similarity), a hybrid metric formula is used to find the optimal matching index. : , in, Hyperparameters for balancing weights (e.g., taking...) =0.5), express Norm.

[0034] Based on the best index retrieved , to transform the original continuous features Replace with the corresponding discrete embedding vector Thus, the quantized feature map is obtained. : , During the backpropagation process, due to The operation is not differentiable, so a straight-through estimator (STE) is used to... The gradient is directly copied to ,Right now This enables end-to-end network training.

[0035] To accurately find embedding vectors that match continuous features, a weighted matching algorithm combining Euclidean distance and cosine similarity is adopted. This algorithm considers both the numerical similarity of features and their directional consistency, effectively improving matching accuracy. Its mathematical expression is: , This yields the quantized discrete representation. .in This is a weighting coefficient, which can be dynamically adjusted according to the type of spatiotemporal physical field. It can be appropriately increased for spatiotemporal physical fields with gentle fluctuations (such as sea surface temperature fields). The value can be appropriately reduced for highly dynamic spatiotemporal physical fields (such as typhoon wind fields). Value selection. At the same time, the dynamic discrete latent space has a complete adaptive update mechanism. After each preset training round, the system will automatically calculate the usage rate and reconstruction contribution of each embedding vector, eliminate embedding vectors with a usage rate lower than a set threshold and a low reconstruction contribution, and generate an equal number of new embedding vectors based on the feature distribution of the current training data through a clustering algorithm. This ensures that the discrete latent space can always accurately adapt to the feature changes of the spatiotemporal physical field and avoid the decline in representation ability caused by the solidification of embedding vectors.

[0036] S33. Finally, there is the decoder module. Its core design idea is to gradually recover the hierarchical features of the spatiotemporal physical field during the feature dimension restoration process, avoiding problems such as feature blurring and loss of detail that are prone to occur in traditional decoders. The decoder quantizes the vector... Reconstructed approximations of the original reanalysis spatiotemporal physics prediction input. The mathematical mapping relationship is as follows: , in, The decoding mapping function is composed of a cross-scale deconvolution layer, an adaptive upsampling module, and a residual connection layer. The cross-scale deconvolution layer progressively expands the feature map size to achieve macroscopic feature reconstruction; the adaptive upsampling module dynamically selects an appropriate interpolation algorithm based on the scale characteristics of the input features to ensure smooth transitions between different regions; and the residual connection layer directly introduces features from the encoder's intermediate layers into the corresponding decoder layers, effectively supplementing fine-grained feature information lost during quantization and significantly improving the accuracy of the reconstructed data. Through the synergistic effect of these structures, the discrete latent representation is ultimately... Restore to input Reconstructing data with consistent dimensions This ensures that core information such as the spatiotemporal structure and numerical range of the spatiotemporal physical field is fully preserved during the compression-decompression process.

[0037] When training the self-prediction model, the model internally uses a composite objective function (or loss function) for optimization. This function preferably includes a weighted combination of one or more of the following functional loss terms: The composite objective loss function of the first-stage self-prediction model The definition is as follows: , S331. Wherein, the first term is the prediction target term (or reconstruction loss), used to measure the distortion or dissimilarity between the reconstructed data and the original input data; mean squared error (MSE) is used to measure the input dense physical field. With reconstructing the physical field The difference between them, the forecast loss The calculation formula is as follows: , The parameters are explained as follows: B: Batch Size, representing the number of samples input in a single training run; C: Channel, representing the channel dimension of the physical field feature map; H and W: the height and width of the feature map, respectively. : The square of the L2 norm, used to calculate the sum of squared errors per pixel between the input and the reconstructed physics field.

[0038] S332. The second term is the discrete latent space learning objective term (or quantization loss), which functions to make the embedding vectors in the discrete latent space move closer to their corresponding continuous feature representations. This process can employ specific gradient control strategies (e.g., but not limited to stopping gradients) to stabilize training. Discrete latent space learning objective term Used to update codebook vectors To make its distribution as close as possible to the encoder output features To avoid gradient backpropagation interfering with encoder parameter updates, a stopping gradient operator is introduced here. (Stop-gradient) blocks the gradient flow on the encoder side. The calculation formula is as follows: , Explanation of the meaning of each parameter and operator: The encoder processes the input samples. Output feature representation; The vector elements in the codebook are the basic building blocks of the discrete latent space; : Stop gradient operator, its function is to shield the gradient calculation of tensors within the operator, allowing only the codebook vector to be gradient-calculated. Updated via backpropagation of loss; : The square of the L2 norm, used to calculate the sum of squared errors per pixel between the input and the reconstructed physics field.

[0039] S333. The third term is the encoder commitment objective (or commitment loss), which functions to make the encoder output closer to its quantized discrete representation in the feature space, so as to improve the stability of the encoder output; the weights of each term in the composite objective function can be balanced by one or more hyperparameters.

[0040] The encoder promises target items Its core function is to constrain the output features of the encoder. To ensure stability and prevent frequent fluctuations during training, the encoder output is driven to converge towards the target codebook vector. Gradient backpropagation is used to prevent loss of the codebook vector. The update causes interference, so a stopping gradient operator is applied to the codebook vector here. The gradient flow on the blocking codebook side is calculated using the following formula: , Explanation of the meaning of each parameter and operator: The encoder processes the input samples. Output feature representation; The vector elements in the codebook are the basic building blocks of the discrete latent space; : Stop gradient operator, its function is to shield the gradient calculation of tensors within the operator, allowing only the codebook vector to be gradient-calculated. Updated via backpropagation of loss; : The square of the L2 norm, used to calculate the sum of squared errors per pixel between the input and the reconstructed physics field.

[0041] Reference Figure 3 , S4. Load the pre-trained module into the second-stage model (which can be understood as a predictive forecasting module) and fix the parameters. Train the second-stage model using time-series samples and a composite loss function. The core task of the second stage is to train a system that can map sparse sensor data to the correct index in this pre-trained discrete latent space.

[0042] The second stage is the core execution stage of the algorithm. Its core objective is to achieve high-precision prediction of dense spatiotemporal physical fields based on sparse sensor observation data and leveraging the hierarchical quantization module pre-trained in the first stage. This is achieved through three steps: dynamic modality fusion coding, lightweight hierarchical attention prediction, and hierarchical code-level loss optimization. Throughout this stage, the hierarchical discrete latent space pre-trained in the first stage remains fixed. and layered decoder The parameters are set to train only the newly added dynamic modality fusion encoder and lightweight hierarchical attention network, ensuring that pre-trained features are not disturbed, while improving the adaptability to sparse data. The following are detailed steps, formulas, and a complete explanation of the variable meanings: S41. This step first reuses the pre-training results from the first stage, and then, in response to the heterogeneity of sparse sensor data, extracts high-quality fusion features through a dynamic modal fusion encoder (DMFE) to solve the problem of the original algorithm's single adaptation to sensor data types.

[0043] S411. Load the three-level discrete latent space and hierarchical decoder completed in the first stage of training. The specific content is as follows: Global layer discrete latent space M=128 (size of discrete latent space), embedding vector dimension ; Discrete latent space of the region layer N=256, embedding vector dimension ; Local layer discrete potential space K=512, embedding vector dimension ; Layered decoder It includes a global upsampling branch, a region upsampling branch, and a local upsampling branch. The output dimension is consistent with the original reanalysis spatiotemporal physics data, and the parameters are fixed and do not participate in the training of this stage.

[0044] S412. Dynamic fusion of modal features is achieved by encoding the two modal components separately using a dynamic modal fusion encoder (DMFE) and calculating adaptive weights based on information entropy. The core formula is as follows: S4121. The feature extraction process of the sub-encoder is as follows: , , in, It is a spatiotemporal feature sub-encoder (consisting of 2 convolutional layers + 1 spatiotemporal attention layer). To determine the dimensionality of the single-modal features before fusion, B is the batch size, and H and W are the height and width of the feature map.

[0045] S4122. Adaptive weight calculation, weights (Spatiotemporal modal weights) and (Physical mode weights) are calculated from the information entropy of the two modal components, ensuring that the mode with more information has a higher weight. The formula is as follows: , , in, Information entropy of spatiotemporal modal features, Information entropy of physical modal features Let be the feature probability distribution of the i-th pixel in the spatiotemporal modality (obtained by feature normalization). Similarly, and satisfying + = 1.

[0046] S4123. Dynamic Fusion Output Finally, the fusion features are obtained. This feature simultaneously incorporates the spatiotemporal context of sensor data and its association with physical elements, providing high-quality input for subsequent attention prediction.

[0047] Reference Figure 4 , S42. A lightweight hierarchical attention network (LHAN) is adopted. Through the serial operation of three-level attention blocks (global-region-local), the long-distance dependence of features at different scales is accurately captured while reducing the amount of computation, and finally outputs a three-level discrete index sequence.

[0048] S421. LHAN consists of 1 global attention block + 3 regional attention blocks + 5 local attention blocks. Information is passed between layers through feature concatenation. Each attention block integrates a structure of convolutional dimensionality reduction + self-attention + residual connection, which solves the problems of large computational cost and many redundant parameters in traditional Transformer.

[0049] The core operation of each attention block is convolutional attention ( First, dimensionality reduction is achieved through convolution to reduce the computational cost of self-attention. Then, the attention mechanism is executed. Finally, residual connections are used to preserve the original features. The calculation formulas for attention blocks at each level are unified as follows: , in, For convolutional attention operations, the internal process is: convolutional dimensionality reduction → linear projection to generate Q / K / V → attention calculation → feature recovery. Taking the global attention block as an example, the detailed formula is as follows: S422. Convolution Dimensionality Reduction ,in, For one-dimensional convolution, the feature dimension is reduced from... =128 dropped to =64, reducing the amount of subsequent attention calculations.

[0050] S423. Generate Q / K / V and embed spatiotemporal location information: , in, It is a learnable linear projection matrix; By embedding the spatiotemporal sinusoidal position of the global layer, the spatial coordinates and temporal order of the features are encoded, thus solving the problem that the attention mechanism is insensitive to position.

[0051] S424. Self-Attention Calculation Feature recovery is then performed. The feature dimension is restored to its original value through one-dimensional convolution. =128, to ensure consistency with the input dimension.

[0052] S425. After the attention block operations at each level are completed, the attention features are projected onto the discrete latent space size dimension of the corresponding discrete space through an independent linear output layer, and the index probability distribution is output: , in, These are the final output features for global, regional, and local attention blocks, respectively. These are the index probability distributions for three levels, with M=128, N=256, and K=512 corresponding to the discrete potential space sizes of each level.

[0053] S426. Based on the maximum value index of the probability distribution, retrieve the corresponding embedding vector from the three-level discrete latent space and combine them to form hierarchical quantization features: , , , Will Hierarchical decoder with fixed input parameters It can output the predicted dense spatiotemporal physical field. .

[0054] S427. Regarding the training objective, the sequence prediction module and sensor encoder module are trained to perform predictions while keeping the discrete latent space and decoder parameters fixed. This stage preferably employs one or more code-level losses for optimization, without using reconstruction loss at the image level. The training objective of the second stage is to map sparse sensor features to a pre-trained discrete codebook index. Therefore, a weighted combination of classification loss and feature alignment loss is used to construct the total loss. The specific definition is as follows.

[0055] S4271. The training objective may include a discrete index prediction target term, which functions to measure the classification or ranking difference between the predicted index sequence and the target index sequence, thereby improving the accuracy of the model in predicting discrete indices. Index prediction in the discrete latent space is essentially a classification task. The cross-entropy loss function is used to optimize the index prediction accuracy at three levels: global (G), region (R), and local (L), as shown in the following formula: , The parameters in the formula have the following meanings: The ground truth index generated by the pre-trained model in the first stage has dimensions that match the codebook size of the corresponding level. The predicted probability distribution of the output of a lightweight hierarchical attention network, representing the model's response to the first... Classification confidence of each index position; The total number of discrete indices for a single sample at the corresponding level.

[0056] S4272. The training objective may further include a feature alignment objective term, which functions to measure the similarity or distance between the intermediate feature representations generated from sparse data and the discrete basis vectors corresponding to the target index, thereby forcing low-quality features to approach their corresponding discrete latent features in the feature space, thus reducing the difficulty of predicting discrete latent coding index sequences. To accelerate model convergence and reduce the distribution shift caused by modal differences, a regression loss in the feature space is introduced as an auxiliary constraint, forcing the fused sensor features... The numerical approximation of the target discrete embedding vector combination is as follows: , The parameters in the formula have the following meanings: : Sensor feature vectors obtained by feature extraction and fusion of sparse sensor data; The target quantization feature obtained by looking up the table from the real index is the result of combining the corresponding vectors in the discrete codebook; The stop gradient operator blocks gradient backpropagation to the pre-trained codebook, thus preventing disruption of the discrete latent space distribution learned in the first stage. : The square of the L2 norm, used to measure the Euclidean distance between sensor features and target quantization features.

[0057] By weighting and combining the two losses mentioned above, we obtain the final optimization objective for the second stage: , In the formula This is the regularization coefficient, used to balance the contribution weights of classification loss and feature alignment loss. Based on experimental experience, a value of 0.1 is recommended.

[0058] In this invention, the weighting factors of the combination are hyperparameters that can be adjusted according to training requirements, and their specific values ​​do not constitute a limitation of this invention. The network after this stage exhibits excellent effectiveness and robustness in the task of predicting dense physical fields from sparse ocean sensor observations.

[0059] S5. Based on the validation set, optimize the model parameters, and finally generate high-resolution spatiotemporal physical field predictions through inference from sensor observation samples on the test set, including: After the model is fully trained, the system can be used for practical inference and prediction. Real-time sparse sensor data streams are input into the system, processed by feature extraction and sequence prediction modules, and output as a predicted index sequence. This sequence is then fed into a fixed decoder, enabling rapid and efficient prediction of high-resolution dense spatiotemporal physical fields at current and future times. The specific implementation is as follows: S51. Evaluate on the validation set: Perform inference on the independent validation set using the trained model, and calculate the physics prediction error (such as mean absolute error, root mean square error, etc.). To quantitatively evaluate the model's physics prediction performance, root mean square error (RMSE) and mean absolute error (MAE) are used as the core evaluation metrics. The specific calculation process and metric descriptions are as follows: Assume the test set contains A sample, for any time... : : The actual physical field data at time t, where The height of the physical field feature map. Width; The predicted field data output by the model, with dimensions different from the actual physical field. completely consistent; The location of the nth sample in the physical field ( , The actual value at ) ; The location of the nth sample in the physical field ( , The predicted value at point ) matches the actual value exactly.

[0060] RMSE is calculated by taking the square root of the sum of the squares and mean of the errors between the actual and forecast values. It is highly sensitive to large errors and can effectively measure forecast bias under extreme weather conditions and other scenarios. The calculation formula is as follows: , MAE is the mean of the absolute errors between the actual and forecast values. It is less affected by extreme values ​​and can objectively reflect the overall average deviation level of the forecast results. The calculation formula is as follows: , The system will traverse all samples in the test set, calculating and statistically analyzing the global mean of RMSE and MAE pixel by pixel. Using RMSE as the core criterion, if its value is lower than a preset threshold (for example, for wind speed physical field forecasting tasks, an error threshold of < 1.5 m / s is set), then the model forecasting performance is determined to meet the actual release standard.

[0061] S52. Dynamic parameter adjustment: Based on the performance on the validation set, the model hyperparameters (such as learning rate, convolutional kernel weights) and network structure (such as number of layers, normalization method) are dynamically adjusted through backpropagation algorithm and gradient descent method. S53. Make good use of the early stopping mechanism. If the validation set error does not decrease for N consecutive iterations, terminate the training to avoid overfitting.

[0062] Experimental verification Reference Figure 6 , Figure 7 , To verify the effectiveness of the forecasting method based on dense spatiotemporal physical fields observed at atmospheric and oceanic points proposed in this invention, this embodiment selects a typical wind field forecast example during the landfall of Hurricane Horn (09:00 UTC on August 26, 2024) in the Hawaiian Islands. The aim is to verify the effectiveness of the proposed forecasting method under extreme storm conditions. The figure visually demonstrates the wind field forecasting visualization effects of different methods on the U component, V component, and wind speed, and compares the ERA5 reanalysis wind field data as ground truth. By observing the detailed texture and numerical intensity of the spatial distribution, the forecasting capabilities of each model in handling extreme weather events can be clearly evaluated. Figure 6 The left column shows the ERA5 ground truth, the middle column shows the prediction results from the model of this invention, and the right column shows the mean absolute error (MAE) distribution of both. As can be seen from the figure, the prediction results generated by the model of this invention maintain a high degree of consistency with the ERA5 ground truth in spatial structure, regardless of whether it is the U component or the V component. The model successfully captures complex nonlinear hydrodynamic features in the wind field, such as high and low pressure centers and fronts, without exhibiting the over-smoothing or edge blurring phenomena common in traditional interpolation methods. Figure 6 As can be seen from the mean absolute error plot on the right, the error values ​​in most areas are at a very low level (dark blue area). Only in local areas with extremely large wind speed gradients (such as near the center of a cyclone) are there a small amount of error fluctuations, but the magnitude of these fluctuations is still within an acceptable range compared to the wind speed itself. This proves that the vector quantization representation learning method proposed in this invention can effectively encode and recover complex vector field features.

[0063] Reference Figure 7 This demonstrates the model's ability to predict the scalar field of wind speed even with extremely low observation density. For example... Figure 7 As shown in (a) sparse observations, the sensor data points used as model inputs are extremely sparse in spatial distribution (coverage far below 5%), making it difficult to infer the complete physical field shape based solely on visual observation. This intuitively demonstrates the ill-posed problem of recovering surface data from point observations. (Comparison) Figure 7 (b) ERA5 truth value and Figure 7 (c) Forecast wind speed (Ours), the model of this invention relies solely on Figure 7The model successfully reconstructed the complete wind speed field forecast from a small number of discrete points in (a). Of particular note is that for the cyclone / typhoon structure presented in the figure (with obvious eye and spiral rainband features), the model not only restored its location and extent, but also accurately reproduced the high wind speed values ​​in the core area. Figure 7 (d) shows the mean absolute error of wind speed. It can be seen that the error is mainly concentrated at the edge of the core area where wind speed changes are most drastic, while the error in the vast background field is almost zero. (Summary) Figure 6 and Figure 7 The experimental results show that the method proposed in this invention can effectively establish a nonlinear mapping from sparse sensor observations to dense spatiotemporal physical fields. Through the first stage of discrete latent space learning and the second stage of sequence prediction, the model still exhibits superior robustness and prediction accuracy even with extremely sparse data, verifying the feasibility, advancement and robustness of the technical solution of this invention.

[0064] Example 2 Reference Figure 5 This embodiment provides a forecasting system based on dense spatiotemporal physical fields observed at atmospheric and oceanic points, including: The data acquisition module is configured to acquire observation data, including historical dense reanalysis spatiotemporal physical field data and corresponding sparse sensor observation data. The data preprocessing module is configured to perform data preprocessing based on the acquired observation data; The representation learning module is configured to perform vectorized representation learning based on the preprocessed data. The feature extraction module is configured to extract intermediate features based on the representation learning results. The sequence prediction module is configured to perform sequence modeling based on feature extraction data and representation learning results; The field formation module is configured to form a predicted dense spatiotemporal physical field based on extracted intermediate features and embedded spatiotemporal information.

[0065] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for predicting dense spatiotemporal physical fields based on atmospheric and oceanic point observations.

[0066] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions adapted for loading and execution by the processor of the aforementioned prediction method based on dense spatiotemporal physical fields observed from atmospheric and oceanic points.

[0067] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting dense spatiotemporal physical fields based on atmospheric and oceanic point observations, characterized in that, The method comprises the following steps: acquiring observation data, including historical dense reanalysis spatiotemporal physical field data and corresponding sparse sensor observation data; performing data preprocessing based on the acquired observation data; performing vectorization representation learning according to the preprocessed data; extracting intermediate features based on the representation learning result; forming a predicted dense spatiotemporal physical field based on the extracted intermediate features and embedded spatiotemporal information.

2. The method of claim 1, wherein, The data preprocessing based on the obtained observation data includes screening out missing values and abnormal values in sensor observation by using a statistical method and a machine learning algorithm, identifying and correcting invalid data in an incorrect format by a program verification rule, and converting the original format data into a unified format meeting the demand analysis of a mathematical model; then, based on a time alignment principle, sparse sensor data matched therewith is screened out through accurate timestamp matching and interpolation processing, and data standardization is performed to complete data preprocessing and divide the data into a training set, a validation set, and a test set; a training sample pair is constructed, , let the time series be , T is the total number of time steps, and when the time step is taken as a reference, the corresponding sample window data W contains n time steps, i.e. , is the dense reanalysis physical field prediction data at the time step , forming effective historical information accumulation for the state at the time step ; finally, the mean value and the standard deviation of each feature i on the training set are calculated, and each data point is standardized.

3. A method for predicting dense spatiotemporal physical fields based on atmospheric and ocean point observations according to claim 2, characterized in that, The vectorization representation learning according to the preprocessed data comprises mapping input gridded data into a continuous intermediate feature representation by using an encoder, wherein the input tensor shape is reanalysis dense spatiotemporal physical field prediction data, representing a batch size, representing a time series quantity of continuous time resolution, representing a number of multi-element channels, representing a feature map height, representing a feature map width, in order to adapt the model input characteristics, the time series dimension and the channel dimension are combined according to feature correlation weighting to strengthen the feature information of the period of intense change of the spatiotemporal physical field, and finally a new model input tensor is obtained, the reanalysis spatiotemporal physical field prediction data after shape adjustment is input into a multi-scale convolution-attention fusion network MSCA-Net for feature extraction, the MSCA-Net is composed of a plurality of hierarchical feature extraction units in series, each unit comprises a convolution layer, a normalization layer, an activation layer and an attention sublayer, wherein the convolution layer adopts a multi-size convolution kernel to capture different scale features, the normalization layer is used to optimize the feature distribution, the activation layer selects a GELU function to enhance the nonlinear fitting ability of the model, and the attention sublayer calculates the feature response strength, and after a plurality of layers of processing, a continuous representation with multi-scale spatial features is output , and the process is represented as: ， wherein, The mapping function representing the encoder is a composite function composed of the multi-layer feature extraction unit.

4. The method of claim 3, wherein, The vectorization representation learning according to the preprocessed data further comprises mapping continuous feature representation to a discrete latent space by using a vector quantization layer, specifically comprising mapping the continuous feature representation to the discrete latent space by using the vector quantization layer. Adaptive mapping to the discrete latent space to realize discrete expression of the features, the discrete latent space containing a group of embedding vectors M is the size of the discrete latent space, the dimension of the embedding vectors being consistent with the dimension of the continuous feature vectors output by the encoder; in order to accurately find the embedding vectors matching the continuous features, a weighted matching algorithm fusing Euclidean distance and cosine similarity is adopted, which takes into account both the closeness in feature values and the consistency in feature directions, and is expressed as: ， Thus, the quantized discrete representation is obtained wherein are weight coefficients.

5. A method for predicting dense spatiotemporal physical fields based on atmospheric and ocean point observations according to claim 4, characterized in that, The vector quantization learning according to the preprocessed data further comprises using a cross-scale deconvolution-up sampling fusion network CSD-Net of a decoder to gradually recover hierarchical features of the space-time physical field in the process of restoring the feature dimension, and the vector-quantized reconstructed into an approximate value of the original reanalysis space-time physical field prediction input The mathematical mapping relationship is: ， in, The mapping function of the decoder will ultimately convert the discrete latent representation... Restore to input Reconstructing data with consistent dimensions This ensures that the spatiotemporal structure and numerical range information of the spatiotemporal physical field are completely preserved during the compression-decompression process.

6. A method for predicting dense spatiotemporal physical fields based on atmospheric and ocean point observations according to claim 5, characterized in that, The intermediate feature extraction based on the representation learning result comprises loading a three-level discrete latent space and a hierarchical decoder, wherein the global layer discrete latent space , the discrete latent space size M=128, the embedding vector dimension ; the regional layer discrete latent space , N=256, the embedding vector dimension ; the local layer discrete latent space , K=512, the embedding vector dimension ; the hierarchical decoder contains a global upsampling branch, a regional upsampling branch and a local upsampling branch, and the output dimension is consistent with the original reanalyzed spatiotemporal physical field data; two modal components are encoded by a dynamic modal fusion encoder DMF, and adaptive weights are calculated based on information entropy to realize dynamic fusion of modal features.

7. A method for predicting dense spatiotemporal physical fields based on atmospheric and ocean point observations according to claim 6, characterized in that, The two modal components are encoded by the dynamic modal fusion encoder DMF, and adaptive weights are calculated based on information entropy to realize dynamic fusion of modal features. The method comprises the following steps: ， ， wherein, is a spatio-temporal feature sub-encoder, which consists of 2 layers of convolution and 1 layer of spatio-temporal attention, is the dimension of single modality feature before fusion, B is the batch size, H and W are the height and width of the feature map, then the adaptive weight calculation is performed, and the weight spatio-temporal modality weight and The physical modality weight is calculated by the information entropy of the two modality components, which ensures that the modality with more information has a higher proportion, and the formula is: ， ， wherein, is the information entropy of the spatio-temporal modal feature, is the information entropy of the physical modal feature, is the feature probability distribution of the i-th pixel point of the spatio-temporal modal, Similarly, and satisfy = 1; and finally performing dynamic fusion output Finally, the fusion feature is obtained The fusion feature contains the spatio-temporal context of the sensor data and the association with the physical elements, and provides high-quality input for subsequent attention prediction.​ 8. A method for predicting dense spatiotemporal physical fields based on atmospheric and ocean point observations according to claim 7, characterized in that, The intermediate feature extraction based on the characterization learning result further comprises adopting a lightweight hierarchical attention network (LHAN) to capture long-distance dependencies of different scale features through three-level attention block serial operation, and finally output three-level discrete index sequences, wherein the 1-layer global attention block, 3-layer regional attention block and 5-layer local attention block of the LHAN are used to realize information transmission through feature splicing, and the operation of each level of attention block is convolution attention The operation formula of each level of attention block is as follows: wherein, is a convolutional attention operation, followed by a convolutional dimension reduction, reducing the feature dimension from = 128 to = 64, reducing the amount of subsequent attention computation, denoted as: where, is a one-dimensional convolution; finally, the Q, K, and V values are generated and spatiotemporal position information is embedded, denoted as: wherein, is a learnable linear projection matrix; is a global layer spatio-temporal sinusoidal position embedding, encoding the spatial coordinates and temporal order of the features, solving the problem of attention mechanism not being sensitive to position, self-attention computation is represented as: , followed by feature restoration Restoring the feature dimension to = 128, ensuring consistency with the input dimension.

9. A method for predicting dense spatiotemporal physical fields based on atmospheric and ocean point observations according to claim 8, characterized in that, The intermediate features are extracted and spatiotemporal information is embedded to form a predicted dense spatiotemporal physical field. After the operation of each level attention block is completed, the attention features are projected to the discrete latent space size dimension of the corresponding discrete space through an independent linear output layer, and an index probability distribution is output. The method comprises the following steps: ， wherein, are the final output features of the global, regional, and local attention blocks, respectively; are the index probability distributions of the three levels, and M=128, N=256, and K=512 correspond to the sizes of the discrete latent spaces of the three levels, respectively. Then, the maximum index is taken according to the probability distribution, and the corresponding embedding vector is retrieved from the three-level discrete latent space to form a layered quantization feature, which is represented as: ， ， ， will be described below. Layered decoder with fixed input parameters , the predicted dense spatiotemporal physical field can be output .

10. A system for predicting dense spatiotemporal physical fields based on atmospheric and oceanic point observations, characterized in that, The method comprises the following steps: The data acquisition module is configured to acquire observation data, including historical dense reanalysis spatiotemporal physical field data and corresponding sparse sensor observation data; The data preprocessing module is configured to perform data preprocessing based on the acquired observation data; The representation learning module is configured to perform vectorization representation learning according to the preprocessed data; The feature extraction module is configured to extract intermediate features based on the representation learning result; The sequence prediction module is configured to perform sequence modeling based on the feature extraction data and the representation learning result; The field forming module is configured to form a predicted dense spatiotemporal physical field based on the extracted intermediate features and embedded spatiotemporal information.

Images