A dual-path perception fusion correction regional sea surface temperature prediction method

Abstract

The application relates to the cross field of marine science and artificial intelligence, and discloses a regional sea surface temperature prediction method based on dual-path perception fusion correction. Sea surface temperature, sea surface wind, sea surface height and sea surface salinity data are acquired, and the sea surface height and the sea surface salinity constitute marine state data; a convolution long short-term memory branch is used to encode the sea surface wind, the marine state data and the sea surface temperature; a context adaptive message gate is used to construct indirect correction messages from the sea surface wind through the marine state to the sea surface temperature and direct correction messages from the sea surface wind to the sea surface temperature; a dual-path perception fusion gate determines a bounded path coefficient and a shared scaling coefficient according to the consistency, relative amplitude and correction strength of the two-way messages, and corrects the sea surface temperature hidden state according to the same; and a residual self-regression decoder outputs a sea surface temperature prediction field. The application can retain multivariate time series information, adapt to the space-time changes of auxiliary information, suppress fusion disturbance and self-regression error accumulation, and improve the prediction accuracy and stability.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of marine science and artificial intelligence, and in particular to a method for predicting regional sea surface temperature using dual-path perception fusion correction. Background Technology

[0002] Sea surface temperature (SST) is a crucial environmental factor characterizing the ocean's thermal state and air-sea interactions. Its spatiotemporal variations are closely related to processes such as monsoon activity, sea surface wind stress, ocean circulation, upwelling, fronts, and mesoscale eddies. Regional SST forecasts provide gridded foundational data for marine meteorological monitoring, marine disaster early warning, marine ecological environment analysis, fisheries production support, and offshore operation decision-making. Taking complex marginal seas as an example, monsoon-driven factors, shelf-deep-sea exchange, and complex coastlines and island / reef distribution collectively cause significant spatial heterogeneity and temporal variability, placing high demands on the accuracy and stability of multi-day SST forecasts.

[0003] Existing methods for regional sea surface temperature (SST) prediction mainly include numerical models, statistical prediction methods, and data-driven prediction methods. Numerical models obtain prediction results by solving ocean dynamics and thermodynamic equations, which have clear physical meanings, but they usually depend on initial fields, boundary conditions, and parameterization schemes, and require high computational resources. Statistical prediction methods have lower computational overhead, but their ability to express complex nonlinear spatiotemporal relationships is limited. With the continuous accumulation of satellite remote sensing, reanalysis, and ocean observation data, deep learning models such as convolutional long short-term memory networks, graph neural networks, and Transformers are increasingly being used for regional SST prediction.

[0004] To improve forecasting capabilities, some existing technologies input auxiliary variables such as sea surface wind, sea surface height, and sea surface salinity along with sea surface temperature into the prediction network. Existing multivariate fusion methods typically employ direct concatenation of input channels, shared coding, or attention weighting to leverage multi-source information for more accurate sea surface temperature predictions.

[0005] In summary, current research has the following problems: (1) Sea surface wind, sea surface height, sea surface salinity, and sea surface temperature have different physical meanings, numerical scales, and temporal variation characteristics. When using input splicing or shared feature extraction methods, the features of each variable are easily mixed, making it difficult to fully retain the temporal evolution information of each variable and to effectively utilize the complementary relationship between different variables; (2) Variables from different sources differ in numerical scale, spatial distribution, and local effectiveness. Existing methods usually lack selection control for message sources, receiving status, and spatial location when transmitting auxiliary information, which may allow information with weak correlation or local anomalies to participate in sea surface temperature status updates, affecting the accuracy and stability of the prediction results; (3) Multi-day sea surface temperature prediction usually uses an autoregressive approach to gradually generate prediction results. Deviations generated at the previous prediction time will affect subsequent predictions. When the state deviation generated during the multivariate fusion process enters the prediction process, it may gradually propagate and accumulate as the prediction time increases, leading to an increase in prediction error under long prediction time. Summary of the Invention

[0006] A method for predicting regional sea surface temperature using dual-path sensing fusion correction, characterized by the following steps:

[0007] S1. Acquire sea surface temperature, sea surface wind, sea surface height, and sea surface salinity data for the target sea area. Perform time registration, spatial grid registration, and standardization on the sea surface height and sea surface salinity data respectively. Then, stitch the processed sea surface height and sea surface salinity data into ocean state data.

[0008] S2. Sea surface wind data Input the corresponding convolutional long short-term memory branch to obtain the hidden state at the corresponding time step. With cell state Where w represents the sea surface wind branch, Indicates the current moment.

[0009] S3. Conceal the current sea surface wind status. The previous hidden state of ocean state data Input to a context-adaptive message gate calculates a corrected message from sea surface wind to ocean state. The previous hidden state of the ocean state is added to the residual of the corrected message, and then input into the corresponding convolutional long short-term memory branch to obtain the hidden state at the current time step. With cell state .

[0010] S4. Change the current hidden state of the sea surface wind. The hidden state of sea surface temperature at the previous moment Input to the context-adaptive message gate yields a corrected message from sea surface wind to sea surface temperature. The hidden state of the ocean at the current moment; The hidden state of sea surface temperature at the previous moment Input to the context-adaptive message gate yields a corrected message from ocean state to sea surface temperature. .

[0011] S5. Combine the correction messages for the current two paths with the hidden state of the sea surface temperature from the previous moment. The input is fed into a dual-path-aware fusion gate. Based on the consistency, relative magnitude, and correction intensity of the two correction messages, bounded path coefficients and shared scaling coefficients are determined to obtain the fused correction after shared scaling. .

[0012] S6. After injecting the fused correction after shared scaling into the previous time step's sea surface temperature hidden state as a residual, it is input into the corresponding convolutional long short-term memory branch to obtain the hidden state of the current time step's sea surface temperature. With cell state .

[0013] S7. Repeat steps S2 to S6 until the encoding of the historical input sequence of length L is completed, obtaining the hidden state of the sea surface temperature at the end of the historical sequence. With cell state The data is then input into the residual autoregressive decoder to predict sea surface temperature data for the next K days, where L represents the length of the historical input sequence and K represents the total number of predicted days.

[0014] Further, step S2 specifically includes:

[0015] Let the current time be , , and These represent the sea surface wind data, ocean state data, and sea surface temperature data at the current moment, respectively. Representing branches At the present moment The latent and cellular states, where w, o, and s represent the sea surface wind branch, ocean state branch, and sea surface temperature branch, respectively.

[0016] Convolution of long short-term memory units using sea surface winds Update wind field status:

[0017]

[0018] in, and These represent the latent state and cellular state of the sea surface wind branch at the current moment, respectively.

[0019] Furthermore, step S3 is as follows:

[0020] S31. Utilizing the context-adaptive message gate from sea surface wind to ocean state Constructing correction messages

[0021]

[0022] S32. Use correction information to correct the hidden state of the ocean at the previous moment.

[0023]

[0024] S33. Convolution of Long Short-Term Memory Units with Ocean State Encoding current ocean state data:

[0025]

[0026] in, The group normalization operation represents the ocean state branch. This represents the hidden state of the ocean at the previous moment after modification.

[0027] Furthermore, step S4 is as follows:

[0028] S41. Construct an indirect correction message using a context-adaptive message gate from ocean state to sea surface temperature. :

[0029]

[0030] S42. Constructing a direct correction message using a context-adaptive message gate from sea surface wind to sea surface temperature. :

[0031]

[0032] in, A context-adaptive message gate representing ocean state to sea surface temperature. Context-adaptive message gates for sea surface winds to sea surface temperatures.

[0033] Furthermore, the context-adaptive message gate structure described in steps S3 and S4 is as follows:

[0034] For message passing from source branch a to target branch b, the hidden state of the current source branch and the hidden state of the previous target branch are concatenated along the channel dimension to obtain a joint context representation. :

[0035]

[0036] Concat indicates splicing along the channel dimension.

[0037] Candidate messages are generated based on the joint context representation. Departmental control diagram and bounded global coefficients And using the local departmental control diagram and bounded global coefficients to adjust candidate messages, a correction message is generated:

[0038]

[0039]

[0040]

[0041]

[0042] Where Conv represents convolution operation, GELU represents activation function, GN represents group normalization, GAP represents global average pooling, and MLP represents multilayer perceptron. This represents the Sigmoid function. This represents element-wise multiplication. This represents the upper limit of the global coefficient.

[0043] Furthermore, step S5 is as follows:

[0044] S51. The correlation coefficient between the indirect and direct correction messages is used to characterize the consistency of the two correction messages, wherein the correlation coefficient... for:

[0045]

[0046] in, and These represent the vectors formed by subtracting the mean from the indirect correction message and the direct correction message, respectively. Clip represents the truncation function. This represents a positive number to prevent numerical instability.

[0047] S52. Calculate the root mean square magnitude of the corrected message using the following formula:

[0048]

[0049] in, This indicates a correction message for the root mean square magnitude to be calculated. It is the i-th element obtained by expanding a single sample along the channel dimension and the spatial dimension, and N represents the total number of elements contained in a single sample.

[0050] S53. Calculate the root mean square magnitude of the hidden state using the following formula:

[0051]

[0052] in, This indicates a correction message for the root mean square magnitude to be calculated. It is the i-th element obtained by expanding a single sample along the channel dimension and the spatial dimension.

[0053] S54. Calculate the consistency descriptor The calculation formula is expressed as:

[0054]

[0055] in, This indicates the absolute value operation. Represents the consistency response threshold, and ReLU represents the activation function.

[0056] S55. Let... , Based on the root mean square magnitude, magnitude percentage, and magnitude difference of the two correction messages, a basic relative magnitude descriptor is constructed:

[0057]

[0058] S56. Let... , And construct supplementary relative amplitude descriptors:

[0059]

[0060]

[0061] S57. Let , Based on the amplitude ratios of the two correction messages relative to the previous sea surface temperature latent state, and the amplitude ratio of the combined correction messages relative to the previous sea surface temperature latent state, a basic correction intensity descriptor is constructed:

[0062]

[0063] set up , And construct supplementary and corrected intensity descriptions:

[0064]

[0065]

[0066] S58. Concatenate the three sets of descriptive values ​​to obtain the fused descriptive vector:

[0067]

[0068] in, These represent the consistency descriptor, relative magnitude descriptor, and correction intensity descriptor, respectively.

[0069] S59. Input the fused description vector into the routing network to obtain the indirect path coefficients. and direct path coefficient :

[0070]

[0071] in, This represents the routing network, where T is the Softmax temperature parameter. and These are the lower and upper limits of the path coefficient, respectively.

[0072] S510. The indirect correction message and the direct correction message are weighted and summed using the indirect path coefficient and the direct path coefficient to obtain the fused correction. :

[0073]

[0074] S511. Calculate the magnitude ratio of the fusion correction relative to the previous time step of the sea surface temperature latent state. And determine the shared scaling factor. :

[0075]

[0076]

[0077] in, To adjust the intensity for the target, This is the lower limit of the shared scaling factor.

[0078] S512. Obtain the fusion correction after shared scaling by multiplying the fusion correction by the shared scaling factor. .

[0079] Furthermore, step S6 is as follows:

[0080] The shared, scaled fusion correction is injected as a residual into the previous sea surface temperature latent state, and the current sea surface temperature state is updated by encoding:

[0081]

[0082]

[0083] Dropout represents random deactivation. The convolutional long short-term memory unit representing the sea surface temperature branch. This represents the previous hidden state of sea surface temperature after fusion correction, group normalization, and random deactivation.

[0084] Furthermore, step S7 is as follows:

[0085] The residual autoregressive decoder initializes its state with the latent sea surface temperature state and cell state at the end of the historical input sequence, and uses the last observed sea surface temperature field as the initial decoding input. :

[0086]

[0087]

[0088] In the k-th daily prediction step, the decoder state is updated using the decoder's convolutional long short-term memory units:

[0089]

[0090] in, and Let these represent the decoder hidden state and the cell state at the k-th daily prediction step, respectively. This represents the long short-term memory unit of the decoder convolution. This represents the decoding input for the k-th daily prediction step.

[0091] The sea surface temperature increment is generated using a convolutional mapping layer and added to the previous decoded input to obtain the current sea surface temperature prediction field. And use it as the decoding input for the next daily prediction step. :

[0092]

[0093]

[0094] in, This indicates a convolutional mapping layer that maps the decoder's hidden state to the sea surface temperature increment field, where K is the total number of predicted days. During the decoding process, sea surface wind data and auxiliary marine environmental data for future times are not input.

[0095] Compared with the prior art, the present invention has the following beneficial effects:

[0096] Firstly, this invention sets up convolutional long short-term memory branches corresponding to sea surface wind, ocean state, and sea surface temperature, and constructs cross-branch correction messages based on the hidden states of different branches. This enables information transfer between variables while preserving the independent temporal evolution information of each variable, which helps to reduce the mutual mixing of different variable characteristics and make full use of the complementary relationship between multi-source marine meteorological data.

[0097] Secondly, this invention utilizes a context-adaptive message gate to generate correction messages based on the current hidden state of the source branch and the previous hidden state of the receiving branch. It also dynamically adjusts the impact of the two correction messages on the sea surface temperature state through a dual-path sensing fusion gate, enabling the introduction of auxiliary information to adapt to state changes at different times and locations, and reducing the impact of weakly correlated or locally abnormal information on the prediction results.

[0098] Third, this invention determines the path coefficient and shared scaling coefficient based on the consistency, relative amplitude, and correction intensity of the two correction messages, and injects the fused correction after shared scaling into the sea surface temperature latent state in the form of residuals, thereby limiting the excessive disturbance of the sea surface temperature state by the correction information, slowing down the propagation and accumulation of state bias in the autoregressive prediction process, and improving the accuracy and stability of regional sea surface temperature prediction under long prediction time. Attached Figure Description

[0099] Figure 1 This is a schematic diagram of the overall process of the regional sea surface temperature prediction method provided in an embodiment of the present invention.

[0100] Figure 2 This is a schematic diagram of the dual-path sensing fusion correction network structure provided in an embodiment of the present invention.

[0101] Figure 3 This is a schematic diagram comparing the root mean square error of sea surface temperature under different prediction lead times according to an embodiment of the present invention. Detailed Implementation

[0102] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described below with reference to the accompanying drawings. The following description is intended to illustrate possible implementations of this invention, and not to limit the scope of protection of this invention; the following technical features can be combined with each other unless otherwise specified. Equivalent substitutions or modifications made by those skilled in the art without departing from the concept of this invention should fall within the scope of protection of this invention.

[0103] In this embodiment, sea surface temperature is abbreviated as SST, sea surface height as SSH, and sea surface salinity as SSS. SSH and SSS are selected as auxiliary marine environmental elements, and they are combined along the channel dimension to form ocean state data, which is used to provide supplementary information on ocean dynamics and thermohaline state related to SST prediction. Sea surface wind data consists of easterly and northerly wind components. The branches are identified by w, o, and s, which represent the sea surface wind branch, ocean state branch, and sea surface temperature branch, respectively.

[0104] Reference Figure 1 The regional sea surface temperature prediction method provided by this invention includes: acquiring and preprocessing multi-source ocean and meteorological data of the target sea area; performing three-branch coding in the order of sea surface wind, ocean state and sea surface temperature; constructing cross-branch correction messages through context-adaptive message gates; fusing indirect correction messages and direct correction messages through dual-path-aware fusion gates and updating sea surface temperature state; and outputting a multi-day sea surface temperature prediction field of the target sea area through a residual autoregressive decoder.

[0105] In this embodiment, historical SST, SSH, SSS, easterly wind, and northerly wind data for the target sea area are time-aligned, spatially grid-aligned, and normalized. SSH and SSS are combined along the channel dimension to form ocean state data, and an ocean mask is constructed based on the effective ocean grid points. The processing employs conventional methods in ocean data prediction, and training and evaluation are performed only on the effective ocean grid points.

[0106] The historical input sequence is constructed using a sliding time window of length L. The network outputs a predicted SST sequence for the next K days based on the historical input sequence:

[0107]

[0108] in, , , and These represent the sea surface wind data, ocean state data, and sea surface temperature data at the current time τ, respectively. The parameter is A dual-path sensing fusion correction network.

[0109] Sea surface wind data, ocean state data, and SST data are input into their respective convolutional long short-term memory branches. The three branches use conventional convolutional long short-term memory units but have independent network parameters, hidden states, and cell states. The state update of any branch is uniformly denoted as:

[0110]

[0111] in, Indicates branch At the present moment Input, and These represent the hidden state and cell state of the corresponding branch, respectively. This represents the corresponding convolutional long short-term memory unit.

[0112] Cross-branch correction messages only affect the hidden state of the receiving branch and do not directly rewrite the cell state of the receiving branch. This allows each branch to retain the independent temporal memory of the corresponding variable group while receiving external corrections.

[0113] Reference Figure 2 At the current moment The network first updates the sea surface wind branch, then uses the sea surface wind information to correct and update the ocean state branch. Subsequently, the ocean state branch and the sea surface wind branch form indirect and direct paths to the sea surface temperature branch, respectively. After dual-path sensing fusion, the sea surface temperature branch is updated.

[0114] First, sea surface winds are used to convolve long short-term memory units. Update sea surface wind status:

[0115]

[0116] Current sea surface wind concealment status And the previous ocean state hidden state Context-adaptive message gate for inputting sea surface winds to ocean state Received sea surface wind to ocean state correction message And inject it into the hidden state of the previous ocean state in the form of residuals:

[0117]

[0118]

[0119] Among them, for the target branch that receives cross-branch correction messages This represents the group normalization operation corresponding to the target branch x. and These represent the group normalization operations corresponding to the ocean state branch and the sea surface temperature branch, respectively. This represents the hidden state of the ocean at the previous moment after modification.

[0120] Using current ocean state data The hidden state of the ocean at the previous moment after modification. and the previous cell state Update ocean status:

[0121]

[0122] Utilizing context-adaptive message gates from ocean state to sea surface temperature Constructing indirect correction messages :

[0123]

[0124] Context-adaptive message gates for sea surface winds to sea surface temperatures Constructing direct correction messages :

[0125]

[0126] The two messages have different information transmission histories and remain independent of each other before entering the sea surface temperature branch.

[0127] For message passing from source branch a to target branch b, the context-adaptive message gate concatenates the current hidden state of the source branch and the hidden state of the previous target branch along the channel dimension to obtain a joint context representation:

[0128]

[0129] Generate candidate messages that match the target branch channel number based on the joint context representation. :

[0130]

[0131] Where Conv represents convolution operation and GN represents group normalization.

[0132] Generate local control graphs based on joint context representations. It is used to select candidate messages at the channel and spatial grid levels:

[0133]

[0134] Global average pooling is performed on the joint context representation, and bounded global coefficients are generated through a multilayer perceptron:

[0135]

[0136] Where GAP represents Global Average Pooling, and MLP represents Multilayer Perceptron. This represents the upper limit of the global coefficient for the corresponding message passing direction.

[0137] By using local departmental control graphs and bounded global coefficients to adjust candidate messages, cross-branch correction messages are obtained:

[0138]

[0139] in, This represents element-wise multiplication. Local departmental control diagrams enable fine-grained selection at the spatial and channel levels, while bounded global coefficients control the overall strength of the cross-branch correction.

[0140] Each sample and Vectors are formed along the channel dimension and the spatial dimension, and the mean of each vector is subtracted to obtain the centered vector. and The consistency between the two messages is represented as:

[0141]

[0142] in, Represents the dot product of vectors. Let L2 denote the norm, and Clip denote the cutoff function. This represents a positive number to prevent numerical instability.

[0143] The root mean square magnitude of the corrected message is calculated using the following formula:

[0144]

[0145] in, This indicates a correction message for the root mean square magnitude to be calculated. It is the i-th element obtained by expanding a single sample along the channel dimension and the spatial dimension, and N represents the total number of elements contained in a single sample.

[0146] The root mean square magnitude of the hidden state is calculated using the following formula:

[0147]

[0148] in, This indicates a correction message for the root mean square magnitude to be calculated. It is the i-th element obtained by expanding a single sample along the channel dimension and the spatial dimension.

[0149] Consistent descriptor It can be represented as:

[0150]

[0151] in, This represents the consistency response threshold.

[0152] set up , Based on the root mean square magnitude, magnitude ratio, and magnitude difference of the two correction messages, a basic relative magnitude descriptor is constructed:

[0153]

[0154] set up , And construct supplementary relative amplitude descriptors:

[0155]

[0156]

[0157] set up , Based on the amplitude ratios of the two correction messages relative to the previous sea surface temperature latent state, and the amplitude ratio of the combined correction messages relative to the previous sea surface temperature latent state, a basic correction intensity descriptor is constructed:

[0158]

[0159] set up , And construct supplementary and corrected intensity descriptions:

[0160]

[0161]

[0162] The three sets of descriptors are concatenated to obtain the fused descriptor vector:

[0163]

[0164] The fused description vector is input into the two-layer routing network. :

[0165]

[0166] in, , , and Indicates trainable parameters, This represents a non-linear activation function.

[0167] The routing network outputs two path logarithmic values, which are then subjected to temperature scaling, Softmax normalization, factor adjustment, and upper and lower bound truncation to obtain the indirect path coefficients. and direct path coefficient :

[0168]

[0169] in, This represents the Softmax temperature parameter. and These represent the lower and upper bounds of the path coefficient, respectively. These upper and lower bounds are used to prevent a single path from being excessively suppressed or amplified.

[0170] The fused correction is obtained by weighted summation of the indirect and direct correction messages based on the bounded path coefficients of the two paths:

[0171]

[0172] Calculate the magnitude ratio of the fusion correction relative to the previous sea surface temperature latent state:

[0173]

[0174] Calculate the shared scaling factor based on the amplitude ratio:

[0175]

[0176] in, Indicates the target correction strength. This indicates the lower limit of the shared scaling factor.

[0177] Path coefficient and Used to adjust the relative distribution of two messages, sharing a scaling factor. Used to control the overall injection strength for fusion correction. Shared scaling applies simultaneously to the weighted sum of the two messages, without altering the already determined relative relationship between the two path coefficients.

[0178] The fusion correction, after shared scaling, is injected as a residual into the previous sea surface temperature latent state:

[0179]

[0180] Dropout represents random deactivation.

[0181] Update sea surface temperature status using current SST data, the corrected previous hidden state, and the previous cell state:

[0182]

[0183] Where Φs represents the convolutional long short-term memory unit of the sea surface temperature branch, This represents the previous hidden state of sea surface temperature after fusion correction, group normalization, and random deactivation.

[0184] The above sequential encoding, message construction, and perception fusion processes are repeated at all L time points to obtain the latent state of sea surface temperature at the end of the historical sequence. and cell state .

[0185] The residual autoregressive decoder is initialized using the sea surface temperature state at the last moment, and the last observed SST field is used as the initial decoding input:

[0186]

[0187]

[0188] On the k-th prediction day, the decoder is used to convolve long short-term memory units. Update the decoder hidden state and cell state:

[0189]

[0190] Using convolutional mapping layers to store the decoder's hidden state Mapped to sea surface temperature increment field:

[0191]

[0192] The sea surface temperature increment field is added to the previous decoding input to obtain the sea surface temperature prediction field for day k, and this prediction field is used as the decoding input for the next daily prediction step:

[0193]

[0194]

[0195] in, and Let these represent the decoder hidden state and cell state at the k-th daily prediction step, respectively. This represents the decoding input for the k-th daily prediction step; This represents the convolutional mapping layer used to convert the hidden state of the multi-channel decoder into a single-channel sea surface temperature increment field; This represents the predicted sea surface temperature increment for day k. This represents the predicted sea surface temperature field for day k.

[0196] During the decoding process, future sea surface wind and auxiliary marine environmental data are not input. The resulting forecast field is de-standardized and organized into gridded SST forecast data, which includes at least the start time, forecast lead time, latitude and longitude coordinates, SST forecast value, and effective ocean mask information, for use by marine meteorological services.

[0197] Before applying the model, end-to-end training is performed using historical input sequences and corresponding future SST ground truth values, with the mean squared error on the effective ocean grid points used as the loss function. Optimizer, learning rate adjustment, and model saving can be performed using conventional training methods. After training is complete, standardized parameters, ocean mask, model structure parameters, and network weights are fixed.

[0198] This implementation selects sea surface temperature, easterly winds, and northerly winds as input variables, and sea surface height and salinity as supplementary input variables. It uses 10 consecutive days of historical data to predict sea surface temperature for the next 15 days. The method uses 32 hidden state channels in each loop branch, employs the Adam optimizer for training, with an initial learning rate of 0.001, a weight decay coefficient of 0.00001, a batch size of 16, a maximum training epoch of 50, and uses the mean squared error over the effective ocean grid as the training loss. Model parameters are selected based on the validation set loss, and prediction results are calculated on the test set.

[0199] The method of this invention was independently trained five times along with ConvLSTM, PredRNN, SwinLSTM, iTransformer, PatchTST, and ResNet3D under the same input variables, data partitioning, ocean mask, historical window, and prediction time. (Refer to...) Figure 3 The method of the present invention maintains a low root mean square error of SST under different prediction lead times, and its advantages are more obvious under long prediction lead times.

[0200] The mean and standard deviation of the average root mean square error (RMSE) from day 1 to day 15 and the RMSE on day 15 obtained from five independent training iterations were calculated, and the experimental results are expressed as "mean ± standard deviation". The average RMSE of the method of this invention from day 1 to day 15 is 0.5509 ± 0.0047 °C, and the RMSE on day 15 is 0.7038 ± 0.0120 °C, which are 6.80% and 5.72% lower than the best-performing comparative method for the corresponding index, respectively. The above results indicate that this invention, by constructing an indirect correction path from sea surface wind to ocean state to sea surface temperature and a direct correction path from sea surface wind to sea surface temperature, and adjusting the path allocation and overall injection intensity according to the consistency, relative amplitude, and correction intensity of the two correction messages, can reduce the disturbance of sea surface temperature state caused by heterogeneous information fusion and the accumulation of errors in multi-day autoregressive prediction, thereby improving the accuracy and stability of regional sea surface temperature prediction.

[0201] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the present invention. The present invention is not limited to the specific embodiments shown herein, and its scope of protection should be determined by the scope defined in the claims.

Claims

1. A method for predicting regional sea surface temperature using dual-path sensing fusion correction, characterized in that, The steps are as follows: S1. Acquire sea surface temperature, sea surface wind, sea surface height and sea surface salinity data of the target sea area, perform time registration, spatial grid registration and standardization on the sea surface height data and sea surface salinity data respectively, and stitch the processed sea surface height data and sea surface salinity data into ocean state data; S2. Sea surface wind data Input the corresponding convolutional long short-term memory branch to obtain the hidden state at the corresponding time step. With cell state Where w represents the sea surface wind branch, Indicates the current time; S3. Conceal the current sea surface wind status. The previous hidden state of ocean state data Input to a context-adaptive message gate calculates a corrected message from sea surface wind to ocean state. The previous hidden state of the ocean state is added to the residual of the corrected message, and then input into the corresponding convolutional long short-term memory branch to obtain the hidden state at the current time step. With cell state , where o represents the ocean state branch; S4. Change the current hidden state of the sea surface wind. The hidden state of sea surface temperature at the previous moment The input to the context-adaptive message gate yields a corrected message from sea surface wind to sea surface temperature, which is then used as the direct corrected message. The hidden state of the ocean at the current moment; The hidden state of sea surface temperature at the previous moment The input is fed into a context-adaptive message gate to obtain a corrected message from ocean state to sea surface temperature, which serves as an indirect corrected message. , where s represents the sea surface temperature branch; S5. Combine the correction messages for the current two paths with the hidden state of the sea surface temperature from the previous moment. The input is fed into a dual-path-aware fusion gate. Based on the consistency, relative magnitude, and correction intensity of the two correction messages, bounded path coefficients and shared scaling coefficients are determined to obtain the fused correction after shared scaling. ; S6. After injecting the fused correction after shared scaling into the previous time step's sea surface temperature hidden state as a residual, it is input into the corresponding convolutional long short-term memory branch to obtain the hidden state of the current time step's sea surface temperature. With cell state ; S7. Repeat steps S2 to S6 until the encoding of the historical input sequence of length L is completed, obtaining the hidden state of the sea surface temperature at the end of the historical sequence. With cell state The data is then input into the residual autoregressive decoder to predict sea surface temperature data for the next K days, where L represents the length of the historical input sequence and K represents the total number of predicted days.

2. The regional sea surface temperature prediction method according to claim 1, characterized in that, Step S2 specifically includes: Let the current time be , , and These represent the sea surface wind data, ocean state data, and sea surface temperature data at the current moment, respectively. Representing branches At the present moment The latent state and cellular state, where w, o and s represent the sea surface wind branch, the ocean state branch and the sea surface temperature branch, respectively; Convolution of long short-term memory units using sea surface winds Update wind field status: in, and These represent the latent state and cellular state of the sea surface wind branch at the current moment, respectively.

3. The regional sea surface temperature prediction method according to claim 1, characterized in that, The process of step S3 is as follows: S31. Utilizing the context-adaptive message gate from sea surface wind to ocean state Constructing correction messages ; S32. Use correction information to correct the hidden state of the ocean at the previous moment. ; S33. Convolution of Long Short-Term Memory Units with Ocean State Encoding current ocean state data: in, The group normalization operation represents the ocean state branch. This represents the hidden state of the ocean at the previous moment after modification.

4. The regional sea surface temperature prediction method with dual-path sensing fusion correction according to claim 1, characterized in that, The process of step S4 is as follows: S41. Construct an indirect correction message using a context-adaptive message gate from ocean state to sea surface temperature. : S42. Constructing a direct correction message using a context-adaptive message gate from sea surface wind to sea surface temperature. : in, A context-adaptive message gate representing ocean state to sea surface temperature. Context-adaptive message gates for sea surface winds to sea surface temperatures.

5. The regional sea surface temperature prediction method according to claim 1, characterized in that, The context-adaptive message gate structure described in steps 3 and 4 is as follows: For message passing from source branch a to target branch b, the hidden state of the current source branch and the hidden state of the previous target branch are concatenated along the channel dimension to obtain a joint context representation. : Wherein, Concat indicates concatenation along the channel dimension; Candidate messages are generated based on the joint context representation. Departmental control diagram and bounded global coefficients And using the local departmental control diagram and bounded global coefficients to adjust candidate messages, a correction message is generated: Where Conv represents convolution operation, GELU represents activation function, GN represents group normalization, GAP represents global average pooling, and MLP represents multilayer perceptron. This represents the Sigmoid function. This represents element-wise multiplication. This represents the upper limit of the global coefficient.

6. The regional sea surface temperature prediction method according to claim 1, characterized in that, The process of step S5 is as follows: S51. The correlation coefficient between the indirect and direct correction messages is used to characterize the consistency of the two correction messages, wherein the correlation coefficient... for: in, and These represent the vectors formed by subtracting the mean from the indirect correction message and the direct correction message, respectively. Clip represents the truncation function. Represents positive numbers to prevent numerical instability; S52. Calculate the root mean square magnitude of the corrected message using the following formula: in, This indicates a correction message for the root mean square magnitude to be calculated. It is the i-th element obtained by expanding a single sample along the channel dimension and the spatial dimension, and N represents the total number of elements contained in a single sample; S53. Calculate the root mean square magnitude of the hidden state using the following formula: in, This indicates a correction message for the root mean square magnitude to be calculated. It is the i-th element obtained by expanding a single sample along the channel dimension and the spatial dimension; S54. Calculate the consistency descriptor The calculation formula is: in, This indicates the absolute value operation. ReLU represents the consistency response threshold, and ReLU represents the activation function. S55. Let... , Based on the root mean square magnitude, magnitude percentage, and magnitude difference of the two correction messages, a basic relative magnitude descriptor is constructed: S56. Let... , And construct supplementary relative amplitude descriptors: S57. Let , Based on the amplitude ratios of the two correction messages relative to the previous sea surface temperature latent state, and the amplitude ratio of the combined correction messages relative to the previous sea surface temperature latent state, a basic correction intensity descriptor is constructed: set up , And construct supplementary and corrected intensity descriptions: S58. Concatenate the three sets of descriptive values ​​to obtain the fused descriptive vector: in, These represent the consistency descriptor, the relative magnitude descriptor, and the correction strength descriptor, respectively. S59. Input the fused description vector into the routing network to obtain the indirect path coefficients. and direct path coefficient : in, This represents the routing network, where T is the Softmax temperature parameter. and These are the lower and upper limits of the path coefficient, respectively; S510. The indirect correction message and the direct correction message are weighted and summed using the indirect path coefficient and the direct path coefficient to obtain the fused correction. : S511. Calculate the magnitude ratio of the fusion correction relative to the previous time step of the sea surface temperature latent state. And determine the shared scaling factor. : in, To adjust the intensity for the target, To share the lower bound of the scaling factor; S512. Obtain the fusion correction after shared scaling by multiplying the fusion correction by the shared scaling factor. .

7. The regional sea surface temperature prediction method according to claim 1, characterized in that, The process of step S6 is as follows: The shared, scaled fusion correction is injected as a residual into the previous sea surface temperature latent state, and the current sea surface temperature state is updated by encoding: Dropout represents random deactivation. The convolutional long short-term memory unit representing the sea surface temperature branch. This represents the previous hidden state of sea surface temperature after fusion correction, group normalization, and random deactivation.

8. The regional sea surface temperature prediction method according to claim 1, characterized in that, The process of step S7 is as follows: The residual autoregressive decoder initializes its state with the latent sea surface temperature state and cell state at the end of the historical input sequence, and uses the last observed sea surface temperature field as the initial decoding input. : In the k-th daily prediction step, the decoder state is updated using the decoder's convolutional long short-term memory units: in, and Let these represent the decoder hidden state and the cell state at the k-th daily prediction step, respectively. This represents the long short-term memory unit of the decoder convolution. This represents the decoding input for the k-th daily prediction step; The sea surface temperature increment is generated using a convolutional mapping layer and added to the previous decoded input to obtain the current sea surface temperature prediction field. And use it as the decoding input for the next daily prediction step. : in, This indicates a convolutional mapping layer that maps the decoder's hidden state to the sea surface temperature increment field, where K is the total number of predicted days. During the decoding process, sea surface wind data and auxiliary marine environmental data for future times are not input.

Images