Regional ocean surface temperature space-time prediction method and system based on space-time feature fusion

Abstract

The invention provides a spatial-temporal feature fusion-based regional ocean surface temperature spatial-temporal prediction method and a spatial-temporal feature fusion-based regional ocean surface temperature spatial-temporal prediction system, and solves the technical problem of low regional medium-short term prediction precision due to weak ocean spatial-temporal feature extraction capability caused by the influence of ocean-land interaction and a small-scale process on an existing ocean land junction and a coastal region. The method comprises the following steps: preprocessing ocean surface temperature data to obtain preprocessed ocean surface temperature data, and performing gradient and differential data sample division to obtain a training set, a verification set and a test set containing an original time-space sequence, a gradient time-space sequence, a flattening time sequence and a differential time sequence; and constructing and using the training set and the verification set to train a spatial-temporal feature fusion model to obtain a trained spatial-temporal feature fusion model, inputting the test set into the trained spatial-temporal feature fusion model for prediction, and then performing reverse normalization to output a final prediction result. The method can be widely applied to the technical field of ocean surface temperature prediction.

Description

Technical Field

[0001] This application belongs to the field of ocean surface temperature prediction technology, and more specifically, it relates to a spatiotemporal prediction method and system for regional ocean surface temperature based on spatiotemporal feature fusion. Background Technology

[0002] Sea surface temperature (SST) is a key physical parameter in marine environmental monitoring. Accurate short-term (7-15 days) spatiotemporal prediction of SST is crucial for maritime safety, fisheries resource management, weather forecasting, and climate change research. Particularly in the marine-land interface and coastal areas, SST changes directly impact coastal ecosystems and human economic activities; therefore, the need for high-precision short-to-medium-term SST forecasts in these regions is particularly urgent.

[0003] The existing Chinese invention patent application CN202311152424.5 proposes a short-term ocean surface temperature (SST) prediction method based on LSTM-ATT-LSTM. It uses an attention mechanism to classify the influence of relevant features and effectively utilizes feature weights for information flow weighting, thus accelerating convergence, improving the short-term dependency of SST information flow, and increasing the prediction accuracy of SST. However, this patent only considers the influence of temporal features on the prediction method and is only applicable to short-term, single-point time series prediction. When applying this method to medium- to short-term, large-scale sea-land interface and coastal areas, it faces many challenges. Specifically, due to the interaction between the ocean and land, thermal differences and sea-land breeze effects occur in this region, leading to more drastic environmental changes and unstable fluctuations, increasing the difficulty of extracting the spatiotemporal features of ocean surface temperature. Simultaneously, small-scale ocean processes such as waves, eddies, and turbulence within the region also disrupt the distribution of sea surface temperature. Therefore, existing ocean surface temperature prediction methods are affected by ocean-land interactions and small-scale ocean processes, resulting in weak ocean spatiotemporal feature extraction capabilities and low short- to medium-term prediction accuracy in the region. Summary of the Invention

[0004] The purpose of this application is to provide a spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion, so as to solve the technical problem that the existing technology has weak marine spatiotemporal feature extraction capability and low short-term prediction accuracy in the marine-land interface and coastal areas due to the influence of marine-land interaction and small-scale processes.

[0005] To achieve the above objectives, a first aspect of this application provides a spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion, comprising the following steps: The ocean surface temperature data is preprocessed to obtain preprocessed ocean surface temperature data. Gradient and differential data samples are then divided to obtain training, validation and test sets containing the original spatiotemporal sequence, gradient spatiotemporal sequence, flattened time series and differential time series. A spatiotemporal feature fusion model is constructed and trained using training and validation sets to obtain the trained spatiotemporal feature fusion model. The test set is then input into the trained spatiotemporal feature fusion model for prediction, and finally, the prediction result is output after inverse normalization.

[0006] Preferably, the spatiotemporal feature fusion model includes a gradient space feature fusion encoding module, a differential time feature fusion encoding module, and a spatiotemporal feature fusion decoding module; The gradient spatial feature fusion encoding module is used to extract spatial features from the original spatiotemporal sequence and its corresponding gradient spatiotemporal sequence respectively. After extraction, the sequences are spliced ​​together to obtain the fused spatial features. After further spatial feature extraction, the sequences are flattened to obtain the spatial features. The differential time feature fusion encoding module is used to extract time features from the flattened time series and its corresponding differential time series respectively, and then concatenate the extracted features to obtain the fused time features. Time features are then extracted again to obtain the final time features. The spatiotemporal feature fusion decoding module is used to fuse extracted spatial and temporal features, and after fusion, feature decoding is performed to capture temporal dependencies, while the predicted features are mapped to the output space.

[0007] Preferably, the gradient spatial feature fusion encoding module includes, in sequence, a first three-dimensional convolutional layer, a first three-dimensional max pooling layer, a splicing layer, a second three-dimensional convolutional layer, a second three-dimensional max pooling layer, and a flattening layer; The differential temporal feature fusion coding module includes, in sequence, a first long short-term memory network, a splicing layer, and a second long short-term memory network; The spatiotemporal feature fusion decoding module includes, in sequence, an additive fusion layer, a transformation layer, a third long short-term memory network, a fourth long short-term memory network, and a fully connected Dense layer.

[0008] Preferably, the formulas for extracting spatial features from the original spatiotemporal sequence and its corresponding gradient spatiotemporal sequence are as follows: ; ; ; ; In the formula, The spatial features are obtained after the original spatiotemporal sequence has passed through the first three-dimensional convolutional layer. The spatial features are obtained after the original spatiotemporal sequence has passed through the first three-dimensional convolutional layer and the first three-dimensional max pooling layer. The spatial features are obtained by passing the gradient spatiotemporal sequence through the first 3D convolutional layer. The spatial features are obtained by passing the gradient spatiotemporal sequence through the first three-dimensional convolutional layer and the first three-dimensional max pooling layer. This is the first 3D convolution operation. The original spatiotemporal sequence, It is a gradient spatiotemporal sequence. This represents the number of filters in the first 3D convolutional layer. This represents the size of the convolution kernel in the first 3D convolutional layer. This is the first three-dimensional max pooling operation.

[0009] Preferably, the formula for further spatial feature extraction is: ; ; In the formula, The spatial features obtained after the fused spatial features are passed through a second 3D convolutional layer. The spatial features obtained after the fused spatial features are passed through a second 3D convolutional layer and a second 3D max pooling layer. This is the second 3D convolution operation. For the integrated spatial characteristics, This represents the number of filters in the second 3D convolutional layer. This refers to the size of the convolution kernel in the second 3D convolutional layer. This is the second three-dimensional max pooling operation.

[0010] Preferably, the formulas for extracting time features from the flattened time series and its corresponding difference time series are as follows: ; ; In the formula, The time features are obtained by flattening the time series and passing it through the first long short-term memory network. The time features are obtained by passing the difference time series through the first long short-term memory network. For the first long short-term memory network operation, For the hidden layer size, This refers to the flattened time series and its corresponding difference time series.

[0011] Preferably, the formula for re-extracting time is: ; In the formula, As a time feature, For the operation of the second long short-term memory network, For the hidden layer size, This represents the temporal characteristics after fusion.

[0012] Preferably, the formula for feature decoding is as follows: ; ; In the formula, The hidden state after passing through the third Long Short-Term Memory network represents the hidden state of length P days. For the third long short-term memory network operation, Let P be the hidden state of length P days. For the hidden layer size, For predicting features, This is the fourth long short-term memory network operation.

[0013] Preferably, preprocessing refers to reading and storing ocean surface temperature data, completing data, cleaning data, and standardizing data to obtain preprocessed ocean surface temperature data.

[0014] A second aspect of this application provides a spatiotemporal prediction system for regional ocean surface temperature based on spatiotemporal feature fusion, comprising: The data preprocessing module is used to preprocess ocean surface temperature data to obtain preprocessed ocean surface temperature data. The gradient and difference data sample partitioning module is used to partition the preprocessed ocean surface temperature data into gradient and difference data samples to obtain training sets, validation sets, and test sets containing the original spatiotemporal sequence, gradient spatiotemporal sequence, flattened time series, and difference time series. The ocean surface temperature prediction module is used to build and train a spatiotemporal feature fusion model using training and validation sets to obtain the trained spatiotemporal feature fusion model. After inputting the test set into the trained spatiotemporal feature fusion model for prediction, the final prediction result is output after inverse normalization.

[0015] The beneficial effects of this application are as follows: This application proposes a spatiotemporal prediction method and system for regional ocean surface temperature based on spatiotemporal feature fusion. Through multi-dimensional feature extraction and fusion mechanisms, it achieves accurate prediction of ocean surface temperature. Specifically, the ocean surface temperature data is first preprocessed, and then the preprocessed ocean surface temperature data is decomposed into the original spatiotemporal sequence and gradient spatiotemporal sequence containing spatial structure, as well as the flattened time series and difference time series reflecting time changes. Subsequently, the prediction is performed by a trained spatiotemporal feature fusion model, followed by inverse normalization to obtain accurate prediction results.

[0016] Furthermore, in the spatiotemporal feature fusion model, a gradient spatial feature fusion encoding module is selected to deeply mine spatial features and effectively capture the complex temperature gradient changes at the land-sea interface; simultaneously, a differential temporal feature fusion encoding module is used to accurately extract temporal evolution patterns; finally, a spatiotemporal feature fusion decoding module is used to organically combine spatial and temporal features and perform feature decoding to capture temporal dependencies. This multi-level, multi-dimensional feature fusion mechanism significantly enhances the ability to characterize small-scale processes and land-sea interaction effects in coastal areas, thereby effectively overcoming the shortcomings of existing technologies in feature extraction in such areas and greatly improving the accuracy of short- and medium-term predictions of regional sea surface temperature. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram of the overall process of a spatiotemporal prediction method and system for regional ocean surface temperature based on spatiotemporal feature fusion provided in an embodiment of this application; Figure 2 A flowchart illustrating data completion provided in one embodiment of this application; Figure 3 A schematic diagram of a three-dimensional spatiotemporal sequence structure provided in an embodiment of this application; Figure 4 This is a schematic diagram of the gradient spatial feature fusion encoding module in a spatiotemporal feature fusion model provided in an embodiment of this application; Figure 5 This is a schematic diagram of the differential temporal feature fusion coding module in a spatiotemporal feature fusion model provided in an embodiment of this application; Figure 6 This is a schematic diagram of the spatiotemporal feature fusion decoding module in a spatiotemporal feature fusion model provided in an embodiment of this application. Detailed Implementation

[0019] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0020] Please see Figure 1 The first embodiment of this application provides a spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion, comprising: S1: Preprocess the ocean surface temperature data to obtain preprocessed ocean surface temperature data.

[0021] The process involves acquiring ocean surface temperature data, preprocessing the data to obtain preprocessed ocean surface temperature data. Preprocessing includes data reading and storage, data completion, data cleaning, and data standardization to obtain the preprocessed ocean surface temperature data.

[0022] In an optional embodiment, longitude, latitude, and sea surface temperature data with a grid size of 720 rows are extracted from an existing world ocean document database. 1440 columns. Using a dynamic region segmentation method, the entire region is uniformly divided vertically into 12 regions, labeled A to L, and horizontally into 24 regions, labeled 01 to 24, resulting in 12... 24 in total, 288 in total, size 60 The system uses a 60-grid region. Each region has a unique row and column combination number. For example, region D09 represents the ocean, specifically the area between 30°N and 45°N latitude and 120°E and 135°E longitude. This area is selected as the experimental area for the prediction method in this application. Each region is named with a corresponding number and stored in a file containing a spatiotemporal sequence of the regional ocean surface temperature over N consecutive days. Its dimension is denoted as This dimension of the spatiotemporal sequence is used as the standard interface for model input.

[0023] Next, please refer to Figure 2 The stored file containing regional ocean surface temperature data is then augmented to obtain the augmented file.

[0024] Specifically, the data completion steps include: opening the acquired data folder containing regional ocean surface temperature data, determining the total number of files (Total) and initializing the count (Num=0); reading the first data file to determine the starting date (Date), extracting the date (date_day) from each data file sequentially, counting (Num=Num+1), and comparing it with the expected date (Date==date_day); if the file's date matches the expected date and the total number of files (Total==count Num), then proceeding to process the next data file (Date=Date+1); if the file's date does not match the expected date, then determining that the date data for that file is missing, identifying the missing date (Date), performing data completion operations (such as re-acquiring or copying the data), updating the total number of files (Total=Total+1), and reading the next data file (Date=Date+1); this process is repeated until all data for the expected dates has been checked and completed.

[0025] Next, the file with the completed data is cleaned to obtain the cleaned file.

[0026] Specifically, data cleaning primarily addresses invalid and outlier values ​​in the completed data file. Invalid values ​​represent NaN values ​​for land areas, which are reassigned to 0. Outliers refer to values ​​outside the reasonable range. Equatorial seas have higher sea surface temperatures due to intense solar radiation, ranging from 28 to 35 degrees Celsius; while polar seas have consistently lower temperatures, typically between -2 and 5 degrees Celsius. This application sets the sea surface temperature threshold range to -2 to 35 degrees Celsius, classifying temperatures outside this range as outliers. The data is cleaned using this threshold range across consecutive time periods, and data is completed using interpolation, copying, or reassignment to 0. The completed file then updates and stores the sea surface temperature data.

[0027] Finally, the cleaned files were standardized to obtain preprocessed ocean surface temperature data.

[0028] Specifically, data standardization refers to eliminating the order-of-magnitude differences between various marine environmental element data caused by variations in numerical ranges and units of measurement, thereby unifying the data to a similar magnitude, reducing memory usage for deep learning model training, and improving training efficiency. This application linearly maps the original data to ocean surface temperature data with a mean of 0 and a variance of 1 by calculating the difference between each data point and the mean of the original data region, and dividing by the regional standard deviation. The specific formula is as follows: ; In the formula, This refers to standardized ocean surface temperature data, i.e., pre-processed ocean surface temperature data. These are the original observation values ​​of ocean surface temperature data points. This represents the mean of the original ocean surface temperature data. This represents the standard deviation of the original ocean surface temperature data.

[0029] S2: Divide the preprocessed ocean surface temperature data into gradient and differential data samples to obtain training, validation and test sets containing the original spatiotemporal sequence, gradient spatiotemporal sequence, flattened time series and differential time series.

[0030] The preprocessed ocean surface temperature data is divided using a sliding window method to obtain sample data, generating a sample dataset. During sample partitioning, training, validation, and test sets are directly obtained for the original spatiotemporal sequence, gradient spatiotemporal sequence, flattened time series, and difference time series. The sliding window size in this application is specified. The sliding step size S=4. The original spatiotemporal sequence, gradient spatiotemporal sequence, flattened time series, and difference time series of the preprocessed ocean surface temperature data from 1982 to 2017 were used as the training set and validation set, with a ratio of 8:2. The original spatiotemporal sequence, gradient spatiotemporal sequence, flattened time series, and difference time series of the preprocessed ocean surface temperature data from 2018 to 2019 were used as the test set.

[0031] Specifically, the gradient and difference data sample partitioning methods are as follows: Using the sliding window method Figure 3 The preprocessed ocean surface temperature data is divided into sample bars to generate a sample dataset, the shape of which is denoted as . The sample dataset is split into training, validation, and test sets.

[0032] S3: Construct and train the spatiotemporal feature fusion model using the validation set and training set to obtain the trained spatiotemporal feature fusion model.

[0033] A spatiotemporal feature fusion model is constructed. The model is trained using a training set and a validation set containing the original spatiotemporal sequence, gradient spatiotemporal sequence, flattened time series and differential time series. By adjusting the structure of the spatiotemporal feature fusion model, a deep learning model suitable for the marine-land interface and coastal areas is constructed.

[0034] Specifically, to extract spatiotemporal features from regional spatiotemporal sequences and uncover spatiotemporal dependencies, a 3D convolutional layer (3D CNN) and a concatenation layer (Concat) are used to concatenate and fuse the spatial features of the original spatiotemporal sequence and the gradient spatiotemporal sequence. A Long Short-Term Memory (LSTM) network and a concatenation layer (Concat) are used to concatenate and fuse the temporal features of the flattened time series and the difference time series, constructing a dual-encoding model structure. Spatial and temporal features are then fused using an additive method. A spatiotemporal feature fusion model based on gradient space and difference time is proposed, consisting of a gradient space feature fusion encoding module, a difference time feature fusion encoding module, and a spatiotemporal feature fusion decoding module. The structure of the spatiotemporal feature fusion model is as follows: Figure 4 As shown.

[0035] The gradient space feature fusion encoding module includes, in sequence: a first three-dimensional convolutional layer (first Conv 3D), a first three-dimensional max pooling layer (first MaxPooling 3D), a concatenation layer (Concat), a second three-dimensional convolutional layer (second Conv 3D), a second three-dimensional max pooling layer (second MaxPooling 3D), and a flattening layer (Flatten).

[0036] The differential temporal feature fusion encoding module includes, in sequence: a first long short-term memory network (first LSTM), a concatenation layer (Concat), and a second long short-term memory network (second LSTM).

[0037] The spatiotemporal feature fusion decoding module includes, in sequence: an additive fusion layer, a transformation layer, a third long short-term memory network (third LSTM) with 50 units, a fourth long short-term memory network (fourth LSTM) with 100 units, and a fully connected Dense layer.

[0038] The spatiotemporal feature fusion model uses a historical input spatiotemporal sequence of length H days to predict an output spatiotemporal sequence of length P days. The specific process is as follows: The gradient spatial feature fusion encoding module is used to process the original spatiotemporal sequence separately. and its corresponding gradient spatiotemporal sequence Spatial features are extracted, then stitched together to obtain fused spatial features. Spatial features are then extracted again and flattened to obtain the final spatial features.

[0039] Specifically, in an optional embodiment, the original spatiotemporal sequence of ocean surface temperature data with a historical input length of H days is used. and its corresponding gradient spatiotemporal sequence This serves as the input to the gradient space feature fusion encoding module. Among them, Let G be the number of the original sequence and the gradient spatiotemporal sequence, and let G be the gradient. The input is the original spatiotemporal sequence. and its corresponding gradient spatiotemporal sequence The shapes are all represented as ,in, For batch size, The length of the historical driving time is the input. The number of rows of data within the region. This represents the number of columns in the data within the region, with 1 channel.

[0040] Specifically, the original spatiotemporal sequence is extracted simultaneously through a first three-dimensional convolutional layer (first Conv3D) and a first three-dimensional max-pooling layer (first MaxPooling 3D). and its corresponding gradient spatiotemporal sequence Spatial features , The formula is as follows: ; ; ; ; In the formula, The spatial features are obtained after the original spatiotemporal sequence has passed through the first three-dimensional convolutional layer. The spatial features are obtained after the original spatiotemporal sequence has passed through the first three-dimensional convolutional layer and the first three-dimensional max pooling layer. The spatial features are obtained by passing the gradient spatiotemporal sequence through the first 3D convolutional layer. The spatial features are obtained by passing the gradient spatiotemporal sequence through the first three-dimensional convolutional layer and the first three-dimensional max pooling layer. This is the first 3D convolution operation. This represents the number of filters in the first 3D convolutional layer, which is set to 40 in this application. The size of the convolution kernel in the first three-dimensional convolutional layer is, in this application, , This is the first three-dimensional max pooling operation. The pooling window size and step size of the first three-dimensional max pooling layer are 2 in this application.

[0041] At this point, the original spatiotemporal sequence and its corresponding gradient spatiotemporal sequence Spatial features obtained after the first three-dimensional convolutional layer and the first three-dimensional max pooling layer , Data shape is .

[0042] Then, a concatenation layer is used to fuse the spatial features obtained after the original spatiotemporal sequence has passed through the first 3D convolutional layer and the first 3D max pooling layer. Spatial features obtained by passing the gradient spatiotemporal sequence through the first 3D convolutional layer and the first 3D max pooling layer The fused spatial features are obtained. The number of channels is increased to retain information from multiple levels. Then, a second 3D convolutional layer (second Conv3D) and a second 3D max-pooling layer (second MaxPooling3D) are used to further extract the fused spatial features. The formula is as follows: ; ; In the formula, The spatial features obtained after the fused spatial features are passed through a second 3D convolutional layer. The spatial features obtained after the fused spatial features are passed through a second 3D convolutional layer and a second 3D max pooling layer. This is the second 3D convolution operation. For the integrated spatial characteristics, This represents the number of filters in the second 3D convolutional layer, which is set to 1 in this application. The size of the convolution kernel in the second three-dimensional convolutional layer is specified in this application. , This is the second three-dimensional max pooling operation. The pooling window size and step size of the second three-dimensional max pooling layer are 2 in this application.

[0043] At this point, the fused spatial features are obtained by passing through the second three-dimensional convolutional layer and the second three-dimensional max pooling layer. The data shape is .

[0044] Finally, the spatial features obtained after the input has passed through the second 3D convolutional layer and the second 3D max pooling layer are flattened. A multidimensional tensor is transformed into a one-dimensional tensor, denoted as spatial feature. Its data shape is .

[0045] The differential temporal feature fusion encoding module respectively processes the flattened time series. and its corresponding difference time series Temporal features are extracted, then concatenated to obtain fused temporal features. Temporal features are then extracted again to obtain the final temporal features.

[0046] Specifically, in an optional embodiment, the H-day ocean surface temperature data includes a flat time series. and its corresponding difference time series Where D is the difference. The flattened time series... and its corresponding difference time series As input to the differential temporal feature fusion coding module, the input data shape is... .

[0047] Specifically, firstly, a flattened time series is extracted simultaneously through a first long short-term memory network (first LSTM). and its corresponding difference time series Temporal features obtained after passing through the first long short-term memory network , The formula is as follows: ; ; In the formula, The time features are obtained by flattening the time series and passing it through the first long short-term memory network. The time features are obtained by passing the difference time series through the first long short-term memory network. For the first long short-term memory network operation, The hidden layer size is 100 in this application.

[0048] At this point, the time series is flattened. and its corresponding difference time series Temporal characteristics obtained after passing through the first long short-term memory network , The data shapes are all .

[0049] Then, a concat layer is used to fuse the flattened time series. and its corresponding difference time series The temporal features obtained after passing through the first long short-term memory network , The fused time characteristics were obtained. Similarly, the number of channels is increased to retain information from multiple levels. Finally, a second Long Short-Term Memory (LSTM) network is used to extract temporal features. The formula is as follows: ; In the formula, As a time feature, For the operation of the second long short-term memory network, For the hidden layer size, this application is , This represents the temporal characteristics after fusion.

[0050] At this time, time characteristics The data shape is .

[0051] The spatiotemporal feature fusion decoding module employs an additive fusion layer to fuse spatial features extracted from a dual-encoding structure. and time characteristics The system uses the third and fourth long short-term memory networks (third LSTM and fourth LSTM) to decode features and capture temporal dependencies, while mapping the predicted features to the output space.

[0052] Specifically, firstly, the spatial features extracted by the gradient spatial feature fusion encoding module and the difference temporal feature fusion encoding module are fused using a feature addition fusion layer. and time characteristics The spatiotemporal hidden features were obtained after fusion. While keeping the number of channels constant, the information content of each dimension of the feature is increased. Then, a transformation layer is used to fuse the spatiotemporal hidden features. Copy and transform to generate the hidden state with a predicted output length of P days. Each hidden state has There are one hidden layer, and the data shape is as follows: .

[0053] Then, the hidden state of length P days Two long short-term memory networks (third LSTM and fourth LSTM) are input for feature decoding to obtain predicted features. The formula is as follows: ; ; In the formula, The hidden state after passing through the third Long Short-Term Memory network represents the hidden state of length P days. For the third long short-term memory network operation, Let P be the hidden state of length P days. For predicting features, This is for the operation of the fourth long short-term memory network, where the hidden layer size of the third long short-term memory network is... The hidden layer size of the fourth Long Short-Term Memory network is 50. It is 100.

[0054] Finally, a fully connected Dense layer is used to predict the features. A nonlinear transformation is performed, mapping the data to the output space. The hidden layer size of the fully connected Dense layer is set to 1. Therefore, through spatiotemporal feature fusion decoding, a spatiotemporal sequence of ocean surface temperature with a predicted output length of P days is obtained, and its data shape is as follows: .

[0055] S4: After the test set is input into the trained spatiotemporal feature fusion model for prediction, the final prediction result is output after inverse normalization.

[0056] The test set of four subsequences (original spatiotemporal sequence, gradient spatiotemporal sequence, flattened time series, and difference time series) is input into the trained spatiotemporal feature fusion model for prediction. After inverse normalization, the final prediction result is obtained. The inverse normalization formula is as follows: ; In the formula, For the final prediction result, The test set consists of four subsequences. These represent the mean and standard deviation of the original ocean surface temperature data, respectively.

[0057] Batch size during spatiotemporal feature fusion model prediction The ocean surface temperature data for day H is in the shape of Flatten it to generate a flattened time series, the data shape of which is The original spatiotemporal sequence is generated by copying B times, and its data shape is... Simultaneously, through the gradient and difference data sample partitioning module, the corresponding gradient spatiotemporal sequence and difference time series are obtained. The data shape of the prediction output of the spatiotemporal feature fusion model is as follows: ,Right now After inverse normalization deformation, the true three-dimensional data of ocean surface temperature for day P (the final prediction result) is obtained, and its data shape is as follows: .

[0058] Please see Figure 1 The second embodiment of this application provides a spatiotemporal prediction system for regional ocean surface temperature based on spatiotemporal feature fusion, comprising: The data preprocessing module is used to preprocess ocean surface temperature data to obtain preprocessed ocean surface temperature data. The gradient and difference data sample partitioning module is used to partition the preprocessed ocean surface temperature data into gradient and difference data samples, resulting in training sets, validation sets, and test sets containing the original spatiotemporal sequence, gradient spatiotemporal sequence, flattened time series, and difference time series. The ocean surface temperature prediction module is used to build and train a spatiotemporal feature fusion model using a validation set and a training set. The trained spatiotemporal feature fusion model is then used to make predictions by inputting the test set into the trained spatiotemporal feature fusion model, and finally, after inverse normalization, the final prediction result is output.

[0059] Example 1: Comparative Experiment To verify the predictive performance of the spatiotemporal feature fusion model, this application is compared with traditional deep learning methods and deep learning combination methods applied to the field of marine prediction. The predictive performance of the spatiotemporal feature fusion model is verified by the root mean square error and prediction accuracy.

[0060] Specifically, the experimental setup in this application is based on a marine surface temperature dataset using satellite remote sensing data, with a spatial resolution of 0.25°C. The experimental area, defined as the sea area between 30° and 45° North latitude and 120° and 135° East longitude, spanning from September 1, 1981 to October 12, 2019, with a daily update resolution, covers the marine-land boundary and coastal region. Data from 1982 to 2017 were used as the training and validation sets, with a ratio of 8:2, for training the regional spatiotemporal sequence model. Data from 2018 to 2019 was used as the test set to evaluate the model's predictive performance, thus completing the regional prediction experiment and research on ocean surface temperature. The experimental data for this application can be found at https: / / www.ncei.noaa.gov / thredds / catalog / OisstBase / NetCDF / V2.0 / AVHRR / catalog.html.

[0061] The experimental area was gridded with a row count equal to the column count of the region (col = 60). The parameters for training the spatiotemporal feature fusion model were set as follows: the Adam algorithm was used for optimization, the learning rate (lr) was 0.001, the batch size (B) was 10, the number of epochs was 100, the mean squared error loss (MSE) was used as the loss function, and cosine similarity was used for evaluation. The experiment involved 12 prediction experiments at 12 time points within a consecutive 12-month period, and the average results of the 12 prediction experiments were analyzed. The comparison results are shown in Table 1.

[0062] Table 1: Comparison of Root Mean Square Error (RMSE) and Prediction Accuracy (PACC) between this application and other methods

[0063] As shown in Table 1, the spatiotemporal feature fusion model of this application has the lowest root mean square error and the highest prediction accuracy compared to traditional methods. Therefore, this application represents a significant improvement in both error and prediction accuracy compared to other prediction models. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0064] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion, characterized in that, Includes the following steps: The ocean surface temperature data is preprocessed to obtain preprocessed ocean surface temperature data. Gradient and differential data samples are then divided to obtain training, validation and test sets containing the original spatiotemporal sequence, gradient spatiotemporal sequence, flattened time series and differential time series. A spatiotemporal feature fusion model is constructed and trained using the training set and the validation set to obtain the trained spatiotemporal feature fusion model. The test set is then input into the trained spatiotemporal feature fusion model for prediction, and finally, the prediction result is output after inverse normalization.

2. The spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion as described in claim 1, characterized in that, The spatiotemporal feature fusion model includes a gradient space feature fusion encoding module, a differential time feature fusion encoding module, and a spatiotemporal feature fusion decoding module. The gradient spatial feature fusion encoding module is used to extract spatial features from the original spatiotemporal sequence and its corresponding gradient spatiotemporal sequence respectively, and then splice them together to obtain the fused spatial features. After extracting spatial features again, the modules are flattened to obtain the spatial features. The differential time feature fusion encoding module is used to extract time features from the flattened time series and its corresponding differential time series respectively, and then splice them together to obtain the fused time features. Time features are then extracted again to obtain the time features. The spatiotemporal feature fusion decoding module is used to fuse the extracted spatial and temporal features, perform feature decoding after fusion to capture temporal dependencies, and simultaneously map the predicted features to the output space.

3. The spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion as described in claim 2, characterized in that, The gradient spatial feature fusion encoding module sequentially includes a first three-dimensional convolutional layer, a first three-dimensional max pooling layer, a splicing layer, a second three-dimensional convolutional layer, a second three-dimensional max pooling layer, and a flattening layer; The differential temporal feature fusion coding module includes, in sequence, a first long short-term memory network, a splicing layer, and a second long short-term memory network; The spatiotemporal feature fusion decoding module includes, in sequence, an additive fusion layer, a transformation layer, a third long short-term memory network, a fourth long short-term memory network, and a fully connected Dense layer.

4. The spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion as described in claim 3, characterized in that, The formulas for extracting spatial features from the original spatiotemporal sequence and its corresponding gradient spatiotemporal sequence are as follows: ； ； ； ； In the formula, The spatial features are obtained after the original spatiotemporal sequence has passed through the first three-dimensional convolutional layer. The spatial features are obtained after the original spatiotemporal sequence has passed through the first three-dimensional convolutional layer and the first three-dimensional max pooling layer. The spatial features are obtained by passing the gradient spatiotemporal sequence through the first 3D convolutional layer. The spatial features are obtained by passing the gradient spatiotemporal sequence through the first three-dimensional convolutional layer and the first three-dimensional max pooling layer. This is the first 3D convolution operation. The original spatiotemporal sequence, It is a gradient spatiotemporal sequence. This represents the number of filters in the first 3D convolutional layer. This represents the size of the convolution kernel in the first 3D convolutional layer. This is the first three-dimensional max pooling operation.

5. The spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion as described in claim 3, characterized in that, The formula for extracting the spatial features again is: ； ； In the formula, The spatial features obtained after the fused spatial features are passed through a second 3D convolutional layer. The spatial features obtained after the fused spatial features are passed through a second 3D convolutional layer and a second 3D max pooling layer. This is the second 3D convolution operation. For the integrated spatial characteristics, This represents the number of filters in the second 3D convolutional layer. This refers to the size of the convolution kernel in the second 3D convolutional layer. This is the second three-dimensional max pooling operation.

6. The spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion as described in claim 3, characterized in that, The formulas for extracting time features from the flattened time series and its corresponding differenced time series are as follows: ； ； In the formula, The time features are obtained by flattening the time series and passing it through the first long short-term memory network. The time features are obtained by passing the difference time series through the first long short-term memory network. For the first long short-term memory network operation, For the hidden layer size, This refers to the flattened time series and its corresponding difference time series.

7. The spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion as described in claim 3, characterized in that, The formula for performing the time extraction again is: ； In the formula, As a time feature, For the operation of the second long short-term memory network, For the hidden layer size, This represents the temporal characteristics after fusion.

8. The spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion as described in claim 3, characterized in that, The formula for performing the feature decoding is as follows: ； ； In the formula, The hidden state after passing through the third Long Short-Term Memory network represents the hidden state of length P days. For the third long short-term memory network operation, Let P be the hidden state of length P days. For the hidden layer size, For predicting features, This is the fourth long short-term memory network operation.

9. The spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion as described in claim 1, characterized in that, The preprocessing refers to reading and storing the ocean surface temperature data, completing the data, cleaning the data, and standardizing the data to obtain the preprocessed ocean surface temperature data.

10. A spatiotemporal prediction system for regional ocean surface temperature based on spatiotemporal feature fusion, applied to the spatiotemporal prediction method for regional ocean surface temperature based on spatiotemporal feature fusion as described in any one of claims 1-9, characterized in that, include: The data preprocessing module is used to preprocess ocean surface temperature data to obtain preprocessed ocean surface temperature data. The gradient and difference data sample partitioning module is used to partition the preprocessed ocean surface temperature data into gradient and difference data samples to obtain a training set, a validation set, and a test set containing the original spatiotemporal sequence, gradient spatiotemporal sequence, flattened time series, and difference time series. The ocean surface temperature prediction module is used to construct and train a spatiotemporal feature fusion model using the training set and the validation set to obtain the trained spatiotemporal feature fusion model. After inputting the test set into the trained spatiotemporal feature fusion model for prediction, the final prediction result is output after inverse normalization.

Images