Marine environment short-term prediction method and system based on conditional adversarial network

Abstract

The invention discloses a marine environment short-term prediction method and system based on a conditional adversarial network, and particularly relates to the field of marine environment short-term prediction, and the method comprises the steps: S1, obtaining multi-dimensional original data covering a marine physical environment and a remote sensing field; s2, performing space-time registration and abnormal value elimination on the multi-dimensional original data, and filling missing values by adopting a moving average method to obtain standardized data; s3, fusing the standardized data into coupling features reflecting the ocean dynamic process, and constructing spatial-temporal feature codes; s4, constructing a conditional adversarial network model based on the spatio-temporal feature codes, driving the conditional adversarial network model to learn a real data distribution rule through a loss function, generating prediction features meeting spatio-temporal conditions, and completing training; according to the method, the three-dimensional spatial-temporal feature tensor reflecting the ocean dynamic process is obtained through multi-source data fusion and coupling feature engineering, and the characterization capability of data and the description precision of a physical mechanism are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of short-term prediction of the marine environment. More specifically, the present invention relates to a short-term prediction method and system for the marine environment based on a conditional adversarial network. Background Art

[0002] In the field of marine environment monitoring and research, short-term prediction is crucial for marine resource development, navigation safety assurance, and ecological environment protection.

[0003] However, the complex marine environment has characteristics of strong non-linearity, multi-parameter coupling, and spatio-temporal dynamic changes. Traditional prediction methods such as physical models and statistical models face significant challenges. Physical models rely on a large number of prior assumptions and simplified equations, and it is difficult to accurately depict fine dynamic processes such as the variation of the thermohaline layer and the propagation of internal solitary waves. Statistical models are restricted by linear assumptions and have insufficient ability to capture the non-linear correlation of ocean current shear and microwave radiation characteristics, resulting in limited prediction accuracy. At the same time, models based on a single data source cannot fully utilize multi-dimensional observation information and lack adaptability to spatio-temporal scale differences, and the risk of prediction failure is high under extreme conditions.

[0004] Therefore, there is an urgent need for a short-term prediction method and system for the marine environment based on a conditional adversarial network. By constructing a generative adversarial framework for multi-scale feature fusion, it is possible to achieve joint prediction of ocean temperature, salinity, flow field, and remote sensing field, solve the bottleneck problems of traditional models in non-linear modeling and multi-source data fusion, and improve the spatio-temporal resolution and accuracy of short-term prediction. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a short-term prediction method and system for the marine environment based on a conditional adversarial network, through the following solutions to solve the problems raised in the above background art.

[0006] To achieve the above object, the present invention provides the following technical solution: A short-term prediction method for the marine environment based on a conditional adversarial network, comprising: S1. Obtain multi-dimensional original data covering the marine physical environment and remote sensing field; S2. Perform spatio-temporal registration and outlier removal on the multi-dimensional original data, and use the moving average method to fill in the missing values to obtain standardized data; S3. Fuse the standardized data into coupling features reflecting the marine dynamic process and construct a spatio-temporal feature encoding; S4. Based on the spatio-temporal feature encoding, construct a conditional adversarial network model, and drive the conditional adversarial network model to learn the distribution law of real data through a loss function, generate prediction features that meet the spatio-temporal conditions, and complete the training; S5. Use the trained model to generate short-term prediction results, and quantitatively evaluate the prediction accuracy through an independent dataset to verify whether the application requirements are met.

[0007] Preferably, the multi-dimensional raw data includes: the vertical temperature gradient of seawater collected by a CTD (Conductivity Temperature Depth profiler) in layers and the coefficient of variation of the thermocline depth ; the standard deviation of the surface seawater salinity obtained by remote sensing inversion combined with buoy measurements ; the vertical shear strength Uz of the geostrophic current measured by an Acoustic Doppler Current Profiler (ADCP); the propagation speed Vw of internal solitary waves monitored by a Synthetic Aperture Radar (SAR); the polarization difference Td of microwave brightness temperature collected by an Advanced Microwave Scanning Radiometer 2 (AMSR2), and the radial velocity shear SC of the ocean current obtained by a High Frequency Ground Wave Radar Network (HFGWRN).

[0008] Preferably, for the spatio-temporal registration: unify the data from different devices to the WGS84 geographic coordinate system, align the timestamps to the UTC second level, and unify the spatial grid resolution to 1 km × 1 km; for the outlier removal: under the assumption of a normal distribution using the 3σ principle, data exceeding the mean ± 3 times the standard deviation are regarded as outliers, and obvious error values are removed in combination with physical constraints; for filling missing values using the moving average method: for short-time series with missing values ≤ 3 sampling points, cubic spline interpolation is used, and for long-time series with missing values, the mean value of historical data in the same period is used for filling.

[0009] Preferably, the coupled features include a thermohaline dynamic coupling feature model and a flow field-remote sensing feature fusion model; the thermohaline dynamic coupling feature model: used to import the vertical temperature gradient of seawater , the standard deviation of the surface seawater salinity and the coefficient of variation of the thermocline depth in the standardized data into the thermohaline dynamic coupling feature model to obtain a thermohaline gradient coupling index , where represents the average depth of the thermocline, and represents the reference temperature gradient; the flow field-remote sensing feature fusion model: used to import the polarization difference Td of microwave brightness temperature, the vertical shear strength Uz of the geostrophic current, and the propagation speed Vw of internal solitary waves in the standardized data into the flow field-remote sensing feature fusion model to obtain a microwave polarization flow field response factor , where represents the microwave brightness temperature difference threshold; for constructing the spatio-temporal feature encoding: expand the thermohaline gradient coupling index CI, the microwave polarization flow field response factor RF, and the radial velocity shear SC of the ocean current in the spatio-temporal dimensions to construct a three-dimensional feature tensor , where T represents the length of the time series, represents the spatial grid dimension, and C = 3 represents the number of feature channels.

[0010] Preferably, the conditional adversarial network model includes a generator and a discriminator. The generator uses a multi-layer spatio-temporal convolutional network as the generator G, and the input includes a conditional vector c and a noise vector z. The generator network structure includes: an input layer, a transposed convolutional layer, and a generator output layer. The discriminator D includes dilated convolution branches with different dilation rates to capture local details and regional macroscopic features respectively, specifically: a discriminator input layer, a multi-branch convolution, and a fusion layer.

[0011] Preferably, for the input layer: the input concatenates the conditional vector c and the noise vector, with a dimension of dc + dz, where dc represents the dimension of the conditional vector, used to describe the number of features included in the conditional vector c, and dz represents the dimension of the noise vector, used to describe the features of the noise vector z. The transposed convolutional layer: there are 3 layers, each followed by a ReLU activation function, gradually expanding the feature map size to ; the generator output layer: a convolutional layer generates a predicted feature tensor , and the activation function is Tanh. The discriminator input layer: receives the real ocean environment feature tensor F or the generated ocean environment feature tensor f, with a dimension of ; the multi-branch convolution: the parallel branches use dilated convolutions with dilation rates of 1, 3, and 5 to extract multi-scale features; the fusion layer: fuses the features of each branch through global average pooling and outputs the discrimination probability .

[0012] Preferably, the loss function includes a generator loss function and a discriminator loss function. The generator loss function is specifically expressed as: ; the discriminator loss function is specifically expressed as: ; where, denotes statistical average, F represents the real ocean environment feature tensor, f represents the generated ocean environment feature tensor, pd represents the real data probability distribution, D() represents the discriminator, G() represents the generator, represents the predicted feature tensor output by the generator, z represents a noise vector subject to a standard normal distribution, c represents a conditional vector containing spatio-temporal information, represents the joint probability distribution of noise and conditions, represents a balance coefficient, represents the L1 loss, used to measure the absolute error between the generated features and the real features.

[0013] Preferably, the short-term prediction result is the ocean environment feature tensor for the next 12 - 48 hours. The quantitative evaluation uses the root mean square error, the mean absolute error, and the correlation coefficient. The root mean square error is specifically expressed as: , the mean absolute error is specifically expressed as: , and the correlation coefficient is specifically expressed as: , where N represents the number of validation samples, represents the i-th real ocean environment characteristic tensor, represents the i-th generated ocean environment feature tensor, Represents the mean value of the real ocean environment characteristic tensor.

[0014] Preferably, the marine environment short-term prediction system based on conditional adversarial network includes: Data acquisition module: covers multi-dimensional raw data of ocean physical environment and remote sensing field; Data processing module: performing spatiotemporal registration and outlier removal on the multi-dimensional raw data, and using a sliding average method to fill in missing values ​​to obtain standardized data; Data fusion module: fusing the standardized data into coupling features reflecting ocean dynamic processes and constructing spatiotemporal feature coding; Model construction and training module: constructing a conditional adversarial network model based on the spatiotemporal feature encoding, driving the conditional adversarial network model to learn the distribution law of real data through a loss function, generating prediction features that meet the spatiotemporal conditions and completing training; Prediction evaluation module: Use the trained model to generate short-term prediction results, and quantitatively evaluate the prediction accuracy through independent data sets to verify whether it meets the application requirements.

[0015] Technical effects and advantages of the present invention: 1. The present invention obtains a three-dimensional spatiotemporal feature tensor reflecting the ocean dynamic process through multi-source data fusion and coupling feature engineering, which solves the problem that the traditional single data source model cannot capture the multi-parameter coupling law, and improves the data characterization ability and the accuracy of the physical mechanism description; 2. The present invention obtains marine environment prediction features that meet spatiotemporal conditions through collaborative training of the conditional adversarial network generator and the multi-scale discriminator, solves the problem of insufficient linear assumptions of traditional physical models and insufficient nonlinear characterization capabilities of statistical models, and achieves the benefit of adaptively capturing the strong nonlinear dynamics of the marine environment; through multi-scale feature fusion and spatiotemporal convolution architecture, prediction results with high spatiotemporal resolution are obtained, which solves the problem of insufficient adaptability of traditional models to different spatial scales and improves the spatiotemporal consistency of prediction results; 3. The present invention obtains the prediction accuracy of the model through quantitative evaluation through independent data set verification and multi-dimensional evaluation indicators, which solves the problem of doubtful prediction reliability caused by the lack of an objective verification system in traditional methods, and achieves the benefit of ensuring the practical application value of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the structure of the method of the present invention; Figure 2Schematic diagram of the system structure of the present invention. Specific embodiments

[0017] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0018] As shown in the attached Figure 1 The short-term ocean environment prediction method based on a conditional adversarial network includes: S1. Obtain multi-dimensional raw data covering the ocean physical environment and the remote sensing field; In this embodiment, specifically, it should be noted that the multi-dimensional raw data includes: the vertical temperature gradient of seawater collected by a CTD (Conductivity Temperature Depth profiler) and the coefficient of variation of the thermocline depth ; the standard deviation of the surface seawater salinity obtained by remote sensing inversion combined with buoy measurements ; the vertical shear strength Uz of the geostrophic current measured by an acoustic Doppler current profiler; the propagation speed Vw of internal solitary waves monitored by a synthetic aperture radar; the microwave brightness temperature polarization difference Td collected by an AMSR2 microwave radiometer, and the sea current radial velocity shear SC obtained by a high-frequency ground wave radar network.

[0019] S2. Perform spatio-temporal registration and outlier removal on the multi-dimensional raw data, and use the moving average method to fill in the missing values to obtain standardized data; In this embodiment, it is specifically necessary to explain that: the spatiotemporal registration: unifying the data of different devices into the WGS84 geographic coordinate system, aligning the timestamps to the UTC second level, and unifying the spatial grid resolution to 1 km × 1 km; the outlier removal: adopting the 3σ principle, that is, under the assumption of normal distribution, data exceeding the mean ±3 times the standard deviation is regarded as abnormal, and the obvious error values ​​are eliminated in combination with physical constraints; the sliding average method is used to fill the missing values: cubic spline interpolation is used for short time series missing ≤3 sampling points, and the long time series missing is filled with the mean of historical data of the same period; it should be understood that the raw data of the marine environment have the following problems: spatiotemporal heterogeneity: the sampling frequency and spatial resolution of different devices are inconsistent, and direct use will lead to feature calculation deviation; noise and outliers: sensor errors, ocean surface white foam interference, etc. may cause the raw data to contain significant outliers; missing values: due to equipment failure or observation conditions, the raw data may have time series or spatial missing. By generating standardized data, reliable input is provided for feature engineering: spatiotemporal registration: unify multi-source data to the same spatiotemporal grid to ensure strict spatiotemporal alignment of parameters during feature calculation; outlier removal: remove erroneous data based on physical constraints (such as the temperature gradient of the thermocline must be negative) and statistical methods (3σ principle) to avoid the derivation of erroneous features (such as the distortion of the coupling index calculated based on the standard deviation of abnormal salinity); missing value filling: fill missing points by interpolation or historical data mean to ensure the continuity of feature calculation (such as the coefficient of variation of thermocline depth requires continuous time series data support). Preprocessing eliminates interference factors to ensure that feature engineering can accurately depict the physical nature of the marine environment and provide a solid data foundation for the efficient training of conditional adversarial networks.

[0020] S3, fusing the standardized data into coupling features reflecting ocean dynamic processes and constructing spatiotemporal feature coding; In this embodiment, it is specifically necessary to explain that: the temperature-salinity dynamic coupling characteristic model is used to convert the vertical temperature gradient of seawater in the standardized data into , Standard deviation of surface seawater salinity and the coefficient of variation of thermocline depth Imported into the temperature-salinity dynamic coupling characteristic model, the temperature-salinity gradient coupling index is obtained ,in, represents the average depth of the thermocline, represents the reference temperature gradient; the flow field-remote sensing feature fusion model is used to import the microwave brightness temperature polarization difference Td, the vertical shear intensity Uz of the geostrophic flow and the internal solitary wave propagation speed Vw in the standardized data into the flow field-remote sensing feature fusion model to obtain the microwave polarization flow field response factor ,in, Represents the microwave brightness temperature difference threshold; the construction of spatiotemporal feature coding: the temperature-salinity gradient coupling index CI, the microwave polarization flow field response factor RF and the ocean current radial velocity shear SC are expanded in spatiotemporal dimensions to construct a three-dimensional feature tensor , where T represents the length of the time series, represents the spatial grid dimension, and C=3 represents the number of feature channels. What needs to be understood is: The synergistic effect of enhanced temperature-salinity gradients. When the two have opposite signs (e.g., warm water has low salinity, cold water has high salinity), the product is negative, indicating strong density stratification; when the signs are the same, the product is positive, which may correspond to a mixed layer or abnormal stratification. The denominator uses square root operation to suppress the dominant role of extreme values, making the index more concerned with the relative change of the gradient rather than the absolute size; The relative strength of the thermocline depth fluctuations is highlighted by a nonlinear transformation (squaring + normalization). When , the term value approaches 0, indicating that the layer is stable; when When , the term value approaches 1, reflecting the violent fluctuation of the stratification; The exponential decay function describes the marginal effect of the temperature gradient. When the temperature gradient exceeds the reference value, the decay accelerates. The direction information of flow field shear is retained. When Uz>0 (the velocity increases with depth), the sign is positive, which may correspond to the upwelling area; when Uz<0, the sign is negative, which may correspond to downwelling or vertical mixing suppressed by the thermocline; The numerator is the product of the microwave difference and the internal wave velocity, which represents the energy synergy effect; the denominator is the sum of the flow field shear and the internal wave velocity, reflecting the competitive contribution of the two to the sea surface roughness. , the term value approaches |Td|, highlighting the modulation of the microwave signal by the internal wave; when the flow field shear is dominant, the term value tends to the flow field energy characteristics; Simulates the nonlinear response of microwave brightness temperature differences. When the difference is less than the threshold, the exponential term approaches 0 and the response factor is suppressed (noise-dominated); when the difference exceeds the threshold, the term value approaches 1 and the response factor increases linearly. Constructs a three-dimensional feature tensor The structure retains the spatiotemporal correlation of the data, making it easier for the subsequent spatiotemporal convolutional network to extract dynamic features.

[0021] S4, constructing a conditional adversarial network model based on the spatiotemporal feature encoding, driving the conditional adversarial network model to learn the distribution law of real data through a loss function, generating prediction features that meet the spatiotemporal conditions and completing training; In this embodiment, it should be specifically noted that: the conditional adversarial network model includes a generator and a discriminator. The generator uses a multi-layer spatio-temporal convolutional network as generator G, and the input includes a conditional vector c and a noise vector z. The generator network structure includes: an input layer, a transposed convolutional layer, and a generator output layer. The discriminator D includes dilated convolution branches with different dilation rates to capture local details and regional macroscopic features respectively, specifically: a discriminator output layer, multi-branch convolution, and a fusion layer. The input layer: concatenates the conditional vector c and the noise vector z, with a dimension of dc + dz. Here, dc represents the dimension of the conditional vector, used to describe the number of features contained in the conditional vector c, and dz represents the dimension of the noise vector, used to describe the features of the noise vector z; The transposed convolutional layer: there are 3 layers, each followed by a ReLU activation function, gradually expanding the size of the feature map to ; The generator output layer: a convolutional layer generates a predicted feature tensor , and the activation function is Tanh; The discriminator output layer: receives the real marine environment feature tensor F or the generated marine environment feature tensor f, with a dimension of ; The multi-branch convolution: the parallel branches use dilated convolutions with dilation rates of 1, 3, and 5 to extract multi-scale features; The fusion layer: fuses the features of each branch through global average pooling and outputs the discrimination probability . The loss function includes a generator loss function and a discriminator loss function; The generator loss function is specifically expressed as: ; The discriminator loss function is specifically expressed as: ; where, represents statistical average, F represents the real marine environment feature tensor, f represents the generated marine environment feature tensor, pd represents the probability distribution of real data, D() represents the discriminator, G() represents the generator, represents the predicted feature tensor output by the generator, z represents a noise vector subject to a standard normal distribution, c represents a conditional vector containing spatio-temporal information, represents the joint probability distribution of noise and conditions, represents the balance coefficient, Denote the L1 loss, which is used to measure the absolute error between the generated features and the real features. It should be understood that the conditional vector c is the prior information used to guide the generator, and usually contains spatio-temporal conditional parameters. For example, the timestamp: used to characterize the time node of the prediction target (such as the encoded value of time units like hours, days, etc.). The spatial coordinates: used to locate the geographical coordinates of the prediction area (such as the normalized values of longitude and latitude); the noise vector z is a random vector that follows the standard normal distribution N(0, 1), which is used to introduce randomness into the generation process, avoid the model generating single-sample, and enhance the diversity of the prediction results; dc is the sum of the dimensions of the above spatio-temporal parameters. For example, if the conditional vector contains "timestamp + longitude + latitude", then dc = 3; dz is a hyperparameter set artificially. For example, dz = 100 means inputting 100-dimensional random noise; through c, the generator is forced to focus on the ocean environmental features at specific spatio-temporal positions (such as the temperature-salinity structure at a certain longitude / latitude in the next 24 hours); through z, the small uncertainties that are not fully observed in the ocean environment (such as high-frequency components like turbulent pulsations and local eddies that are difficult to model) are captured to improve the generalization ability of the prediction results; the transposed convolutional layer: 3 layers: The first layer: a 4×4 convolutional kernel, dilation rate 1, output feature map size 12×50×50, activation function ReLU, which is used to increase the spatial resolution; The second layer: a 4×4 convolutional kernel, dilation rate 2, output feature map size 24×100×100, activation function ReLU, which captures mesoscale spatial features; The third layer: a 4×4 convolutional kernel, dilation rate 1, output feature map size 24×100×100×3, activation function Tanh, which normalizes the values to [-1, 1] to match the real data distribution; The multi-branch convolution: The parallel branches adopt dilated convolutions with dilation rates 1, 3, and 5: Branch 1: dilation rate 1, 3×3 convolutional kernel, which extracts local features (such as the temperature-salinity coupling strength at a single grid point); Branch 2: dilation rate 3, 3×3 convolutional kernel, which extracts regional features (such as the sea current shear trend within a 3×3 grid); Branch 3: dilation rate 5, 3×3 convolutional kernel, which extracts macroscopic features (such as the internal wave propagation direction within a 5×5 grid). The generator loss function: The adversarial loss part: The objective: Force the sample f = G(z, c) output by the generator to be misjudged as real data by the discriminator (i.e., D(f, c) 1)) In ocean prediction, the adversarial loss ensures that the generated features such as temperature-salinity gradients and sea current shears conform to the real distribution law of the target sea area (such as the statistical characteristics of a shallower thermocline in summer); L1 loss: The objective: Directly constrain the numerical difference between the generated data and the real data (such as the absolute error of the temperature-salinity gradient coupling index); Avoid the generator ignoring the actual accuracy of physical quantities; The discriminator loss function: Make the discriminator output = 1 for real data and output for generated data, so as to improve the classification ability of the discriminator and indirectly force the generator to generate higher-quality samples.

[0022] S5. Use the trained model to generate short-term prediction results, and quantitatively evaluate the prediction accuracy through an independent data set to verify whether the application requirements are met.

[0023] In this embodiment, it should be specifically noted that: the short-term prediction result is the ocean environmental feature tensor for the next 12 - 48 hours, and the quantitative evaluation uses the root mean square error, mean absolute error, and correlation coefficient. The root mean square error is specifically expressed as: , and the mean absolute error is specifically expressed as: , and the correlation coefficient is specifically expressed as: , where N represents the number of verification samples, represents the i-th true ocean environmental feature tensor, represents the i-th generated ocean environmental feature tensor, represents the mean value of the true ocean environmental feature tensors. It should be understood that: the root mean square error measures the overall deviation between the predicted value and the true value, assigns a higher weight to larger errors (because the square operation amplifies the influence of outliers), and is suitable for detecting the prediction ability of the model for extreme ocean events (such as sudden internal solitary waves and strong upwelling). If the RMSE increases significantly in extreme samples, it indicates that the model has insufficient modeling of the nonlinear mutation process. The RMSE directly reflects the absolute difference between the predicted value and the true value, which is convenient for business personnel to intuitively understand the error range; the mean absolute error: calculates the average value of the absolute errors between the predicted value and the true value, and is more robust to outliers (avoiding the amplification of individual errors by the square operation), and reflects the prediction deviation level of the model under normal ocean environmental conditions (such as the prediction of sea current speed in a calm sea area). If the MAE is close to the RMSE, it indicates that the data distribution is uniform and there are no significant outliers; the correlation coefficient: measures the linear correlation degree between the predicted value and the true value, with a value range of [0, 1]. The closer the value is to 1, the stronger the trend consistency between the prediction result and the true data, and it is suitable for evaluating the ability of the model to capture the trend of ocean environmental evolution (such as the seasonal change trend of the temperature-salinity gradient). If R² is relatively high, it indicates that the model can effectively depict the spatio-temporal evolution law of physical quantities. The three respectively quantify the model performance from the dimensions of "error range", "average deviation", and "trend correlation", avoiding the one-sidedness of a single indicator. For example, if the RMSE is high but R² is high, it indicates that the model can capture the trend but there are systematic biases, which can be corrected through calibration; if R² is low but the MAE is low, it indicates that the predicted value is close to the mean value but lacks dynamic adaptability, and the spatio-temporal feature extraction mechanism needs to be optimized. By comparing the index changes under different training strategies (such as adjusting coupled features and changing the number of network layers), the model structure is optimized directionally to form a closed loop of "prediction - verification - tuning" to continuously improve the accuracy and reliability of short-term ocean environmental prediction.

[0024] Based on the above solution and attached Figure 2 The present invention also provides a short-term ocean environmental prediction system based on a conditional adversarial network, including: Data acquisition module: Multidimensional raw data covering the ocean physical environment and remote sensing fields; Data processing module: Perform spatio-temporal registration and outlier removal on the multidimensional raw data, and use the moving average method to fill in the missing values to obtain standardized data; Data fusion module: Fuse the standardized data to reflect the coupling characteristics of ocean dynamic processes and construct spatio-temporal feature encodings; Model construction and training module: Construct a conditional adversarial network model based on the spatio-temporal feature encodings, drive the conditional adversarial network model to learn the real data distribution law through the loss function, generate prediction features that meet the spatio-temporal conditions and complete the training; Prediction and evaluation module: Use the trained model to generate short-term prediction results, and quantitatively evaluate the prediction accuracy through an independent data set to verify whether the application requirements are met.

[0025] Secondly: In the accompanying drawings of the disclosed embodiments of the present invention, only the structures related to the disclosed embodiments are involved. Other structures can refer to the general design. Without conflict, the same embodiment and different embodiments of the present invention can be combined with each other; Finally: The above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A short-term prediction method for marine environment based on conditional adversarial network, characterized in that include: S1. Obtain multi-dimensional raw data covering the ocean physical environment and remote sensing field; S2. Performing spatiotemporal registration and outlier removal on the multi-dimensional raw data, and using a sliding average method to fill in missing values ​​to obtain standardized data; S3, fusing the standardized data into coupling features reflecting ocean dynamic processes and constructing spatiotemporal feature coding; the coupling features include a temperature-salinity-dynamic coupling feature model and a flow field-remote sensing feature fusion model; S4, constructing a conditional adversarial network model based on the spatiotemporal feature encoding, driving the conditional adversarial network model to learn the distribution law of real data through a loss function, generating prediction features that meet the spatiotemporal conditions and completing training; S5. Use the trained model to generate short-term prediction results, and quantitatively evaluate the prediction accuracy through independent data sets to verify whether it meets the application requirements.

2. The short-term marine environment prediction method based on conditional adversarial network according to claim 1, characterized in that: The multi-dimensional original data includes: the vertical temperature gradient of seawater collected by a CTD in layers and the coefficient of variation of the thermocline depth ; the standard deviation of the surface seawater salinity obtained by remote sensing inversion combined with buoy measurements ; the vertical shear intensity Uz of the geostrophic current measured by an acoustic Doppler current profiler; the propagation speed Vw of internal solitary waves monitored by a synthetic aperture radar; the polarization difference Td of microwave brightness temperature collected by an AMSR2 microwave radiometer, and the shear SC of the radial velocity of ocean currents obtained by a high-frequency ground wave radar network.

3. The short-term prediction method for marine environment based on conditional adversarial network according to claim 1, characterized in that: The described spatiotemporal alignment: unify the data of different devices into the WGS84 geographic coordinate system, align the timestamps to the UTC second level, and unify the spatial grid resolution to 1 km × 1 km; the described outlier removal: adopt the 3σ principle, that is, under the assumption of normal distribution, data exceeding the mean ±3 times the standard deviation is regarded as abnormal, and obviously erroneous values ​​are eliminated in combination with physical constraints; the described sliding average method is used to fill missing values: for short time series with missing ≤3 sampling points, cubic spline interpolation is used, and for long time series missing, the mean of historical data for the same period is used to fill.

4. The short-term prediction method for marine environment based on conditional adversarial network according to claim 1, wherein: The thermohaline dynamic coupling feature model: is used to import the vertical temperature gradient of seawater in the standardized data , the standard deviation of surface seawater salinity , and the coefficient of variation of the thermocline depth into the thermohaline dynamic coupling feature model to obtain the thermohaline gradient coupling index , where represents the average depth of the thermocline, represents the reference temperature gradient; The flow field-remote sensing feature fusion model: is used to import the microwave brightness temperature polarization difference Td, the geostrophic flow vertical shear intensity Uz, and the internal solitary wave propagation speed Vw in the standardized data into the flow field-remote sensing feature fusion model to obtain the microwave polarization flow field response factor , where represents the microwave brightness temperature difference threshold; The construction of spatio-temporal feature encoding: expands the thermohaline gradient coupling index CI, the microwave polarization flow field response factor RF, and the ocean current radial velocity shear SC in the spatio-temporal dimensions to construct a three-dimensional feature tensor , where T represents the time series length, represents the spatial grid dimension, and C = 3 represents the number of feature channels.

5. The short-term prediction method for marine environment based on conditional adversarial network according to claim 1, characterized in that: The conditional adversarial network model includes a generator and a discriminator. The generator adopts a multi-layer spatiotemporal convolutional network as the generator G, and the input includes a conditional vector c and a noise vector z; the generator network structure includes: an input layer, a transposed convolution layer and a generator output layer; the discriminator D includes hole convolution branches with different expansion rates, which respectively capture local details and regional macro features, specifically: a discriminator input layer, a multi-branch convolution and a fusion layer.

6. The short-term prediction method for marine environment based on conditional adversarial network according to claim 5, characterized in that: The input layer: inputs the concatenated conditional vector c and the noise vector, with a dimension of dc + dz, where dc represents the dimension of the conditional vector, used to describe the number of features contained in the conditional vector c, and dz represents the dimension of the noise vector, used to describe the number of features of the noise vector z; The transposed convolutional layer: consists of 3 layers, each followed by a ReLU activation function, gradually expanding the size of the feature map to ; The generator output layer: a convolutional layer generates a predicted feature tensor The activation function is Tanh; The discriminator input layer: receives the real marine environment feature tensor F or the generated marine environment feature tensor f, with a dimension of ; The multi-branch convolution: the parallel branches use dilated convolutions with dilation rates of 1, 3, and 5 to extract multi-scale features; The fusion layer: fuses the features of each branch through global average pooling and outputs the discrimination probability .

7. The short-term ocean environment prediction method based on conditional adversarial network according to claim 1, characterized in that: The loss function includes a generator loss function and a discriminator loss function; the generator loss function is specifically expressed as: ; the discriminator loss function is specifically expressed as: ; where represents statistical average, F represents the true ocean environment feature tensor, f represents the generated ocean environment feature tensor, pd represents the true data probability distribution, D() represents the discriminator, G() represents the generator, represents the predicted feature tensor output by the generator, z represents the noise vector subject to the standard normal distribution, c represents the conditional vector containing spatio-temporal information, represents the joint probability distribution of the noise and the condition, represents the balance coefficient, represents the L1 loss, which is used to measure the absolute error between the generated features and the true features.

8. The short-term prediction method for marine environment based on conditional adversarial network according to claim 1, wherein: The short-term prediction result is the ocean environmental feature tensor for the next 12 - 48 hours. The quantitative evaluation uses the root mean square error, mean absolute error, and correlation coefficient. The root mean square error is specifically expressed as: , the mean absolute error is specifically expressed as: , the correlation coefficient is specifically expressed as: , where N represents the number of validation samples, represents the i-th true ocean environmental feature tensor, represents the i-th generated ocean environmental feature tensor, represents the mean of the true ocean environmental feature tensors.

9. A short-term prediction system for marine environment based on conditional adversarial network, which is used to implement the short-term prediction method for marine environment based on conditional adversarial network according to any one of claims 1 to 8 above, characterized in that, include: Data acquisition module: covers multi-dimensional raw data of ocean physical environment and remote sensing field; Data processing module: performing spatiotemporal registration and outlier removal on the multi-dimensional raw data, and using a sliding average method to fill in missing values ​​to obtain standardized data; Data fusion module: fusing the standardized data into coupling features reflecting ocean dynamic processes and constructing spatiotemporal feature coding; Model construction and training module: constructing a conditional adversarial network model based on the spatiotemporal feature encoding, driving the conditional adversarial network model to learn the distribution law of real data through a loss function, generating prediction features that meet the spatiotemporal conditions and completing training; Prediction evaluation module: Use the trained model to generate short-term prediction results, and quantitatively evaluate the prediction accuracy through independent data sets to verify whether it meets the application requirements.

Images