A method and system for long-term ocean surface salinity prediction

Abstract

The application belongs to the technical field of microwave remote sensing and detection, and particularly relates to a method and system for long-time sequence marine surface salinity prediction, which obtains input values and forgetting values by inputting monthly salinity data into input gates and forgetting gates, generates a multi-scale periodic coding vector according to a time step, inputs the multi-scale periodic coding vector into a full connection neural network, and performs dimension expansion to obtain a gating tensor; trend features are extracted from a previous cell state; a current hidden state corresponding to the current cell state is restored according to the trend features and the gating tensor to obtain current salinity data; the current salinity data is input into a prediction model for training, the prediction model is modulated according to a loss function, prediction model parameters are obtained when the loss function is minimum, the prediction model is adjusted according to the prediction model parameters, and an optimized prediction model is obtained; and the monthly salinity data set is input into the optimized prediction model to obtain predicted salinity values. The application can improve the accuracy of marine surface salinity prediction without a large number of parameters.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of microwave remote sensing and detection, and particularly relates to a method and system for predicting ocean surface salinity in a long time sequence. BACKGROUND

[0002] In recent years, with the global climate change and the increasing complexity of the marine ecosystem, it is increasingly important to accurately predict the ocean salinity. The ocean surface salinity (SSS) is an important indicator of the interaction between the ocean and the atmosphere, and the salinity change will affect the seawater density, thereby affecting the global ocean circulation and the thermohaline circulation system. By predicting the ocean surface salinity, the climate change can be better predicted. The prediction of the ocean surface salinity has great significance in climate research, extreme weather event prediction and marine ecological protection.

[0003] In the related art, the prediction of the ocean surface salinity mainly adopts a physical model and a statistical model. The physical model is based on the basic principles of oceanography, fluid dynamics and thermodynamics, and predicts the distribution and change of the salinity through a large number of ocean parameter simulations. The statistical model does not directly consider the physical process, but establishes a statistical relationship between the salinity and related variables (such as temperature, wind field, precipitation, etc.) by analyzing historical observation data (such as satellite remote sensing, buoy data, etc.).

[0004] In view of the above related art, for the physical model, a large number of ocean parameters are required when the physical model is constructed, and these parameters differ greatly under different sea areas, depths, temperatures and salinity conditions, and it is difficult to accurately measure. For the statistical model, the relationship is established by finding the rule from the historical data, and the data quality and quantity affect the performance of the model. The data has errors, missing or is not representative, and the model accuracy and generalization ability are poor. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a method and system for predicting the ocean surface salinity in a long time sequence, which can improve the accuracy of the prediction of the ocean surface salinity without a large number of parameters.

[0006] A method for predicting the ocean surface salinity in a long time sequence, comprising:

[0007] obtaining daily brightness temperature data in a preset time period, and preprocessing the daily brightness temperature data to obtain a daily brightness temperature data set;

[0008] inverting the daily brightness temperature data set to obtain corresponding monthly salinity data, and different time monthly salinity data forming a monthly salinity data set, the monthly salinity data including the salinity data corresponding to each geographical grid position in a set regional latitude and longitude range within a month;

[0009] The monthly salinity data set is combined according to time steps to obtain a salinity time sequence, and the salinity time sequence is sequentially input into an input gate and a forget gate according to time steps to obtain an input value and a forget value;

[0010] According to the time step, a multi-scale periodic encoding vector is generated;

[0011] The multi-scale periodic encoding vector is input into a fully connected neural network to obtain a periodic auxiliary gate vector, and the periodic auxiliary gate vector is dimensionally expanded to obtain a gating tensor;

[0012] The previous cell state of the last time step is obtained, and the cell state is used to describe the forgetting degree and the retention degree of the salinity time sequence;

[0013] Trend features are extracted from the previous cell state;

[0014] According to the trend features, the gating tensor, the input value and the forget value, a current cell state is obtained;

[0015] The current cell state is activated to obtain a current hidden state, and the current hidden state is reconstructed to obtain a current salinity data;

[0016] The current salinity data is input into a prediction model to obtain a prediction output value, and according to the prediction output value and a training value, a loss value is calculated, the prediction model is adjusted according to the loss value to obtain a prediction model parameter when the loss function is minimized, and the prediction model is adjusted according to the prediction model parameter to obtain an optimized prediction model;

[0017] Actual monthly salinity data set is input into the optimized prediction model to obtain a predicted salinity value.

[0018] Optionally, the corresponding monthly salinity data is obtained by inverting the daily brightness temperature data set, and the monthly salinity data at different times constitutes a monthly salinity data set, including:

[0019] The daily brightness temperature data is corrected to obtain corrected data;

[0020] The corrected data is subjected to region selection and normalization operation to obtain normalized data;

[0021] According to the normalized data, the corresponding monthly salinity data is obtained, and the monthly salinity data at different times constitutes a monthly salinity data set.

[0022] Optionally, the monthly salinity data set is combined according to time steps to obtain a salinity time sequence, and the salinity time sequence is sequentially input into an input gate and a forget gate according to time steps to obtain an input value and a forget value, including:

[0023] Set the time step;

[0024] combining the monthly salinity data set by time step, a salinity time sequence is obtained;

[0025] obtaining a hidden state of a previous time step;

[0026] inputting the salinity time sequence and the hidden state of the previous time step into an input gate and a forget gate in sequence by time step, obtaining an input value and a forgetting value, and denoted as:

[0027]

[0028]

[0029] wherein, is the input value, denotes the forgetting value, denotes a convolution operation, denotes a Sigmoid activation function, and the output range is [0, 1], , , and are input and historical state pair gating convolution kernels, is the hidden state of the previous time step, and are bias terms, is the salinity time sequence.

[0030] Optionally, the generating a multi-scale periodicity encoding vector according to the time step comprises:

[0031] obtaining a multi-scale periodicity encoding vector formula;

[0032] inputting the time step into the multi-scale periodicity encoding vector formula, obtaining a multi-scale periodicity encoding vector, and the multi-scale periodicity encoding vector formula is denoted as:

[0033]

[0034]

[0035] wherein, is a defined period, k is a scale number, k = 1, 2 or 3, and t represents a difference value of a current year relative to a starting year, is a multi-scale periodicity encoding vector, and the starting year is the earliest year in the monthly salinity data set.

[0036] Optionally, the inputting the multi-scale periodicity encoding vector into a fully connected neural network to obtain a periodicity auxiliary gate vector, and performing dimension expansion on the periodicity auxiliary gate vector to obtain a gating tensor comprises:

[0037] The multi-scale periodic encoding vector is input into a two-layer fully connected neural network to obtain a periodic auxiliary gate vector, and the periodic auxiliary gate vector is obtained, denoted as:

[0038]

[0039] wherein, and is a fully connected network mapping, and is a bias term matched with the current space state, is a Sigmoid activation function, is used for nonlinear transformation, so that the model can more flexibly process the periodic encoding information, is a multi-scale periodic encoding vector.

[0040] The periodic auxiliary gate vector is dimensionally expanded to obtain a gating tensor.

[0041] Optionally, the trend feature extracted from the previous cell state is represented as:

[0042]

[0043] wherein, is a convolution kernel for trend extraction, is a previous cell state, is a hyperbolic tangent activation function, and the output range is [-1, 1], is a trend feature, is a bias term.

[0044] Optionally, the current cell state is obtained according to the trend feature and the gating tensor, and is represented as:

[0045]

[0046] wherein, is a previous cell state, represents a forgetting value, is a periodic response, is an element-wise multiplication, is an input value.

[0047] A system for predicting long-term ocean surface salinity, comprising:

[0048] A first acquisition module is configured to acquire daily skin temperature data in a preset time period, and to preprocess the daily skin temperature data to obtain a daily skin temperature data set.

[0049] The second acquisition module is configured to obtain corresponding monthly salinity data by inversing the daylight temperature dataset, and monthly salinity data at different times form a monthly salinity dataset, wherein the monthly salinity data include salinity data corresponding to each geographical grid position in a set region within a month.

[0050] The first calculation module is configured to combine the monthly salinity dataset according to time steps to obtain a salinity time sequence, and input the salinity time sequence into an input gate and a forget gate in sequence according to time steps to obtain an input value and a forget value.

[0051] The second calculation module is configured to generate a multi-scale periodic encoding vector according to the time steps.

[0052] The gated tensor generation module is configured to input the multi-scale periodic encoding vector into a fully connected neural network to obtain a periodic auxiliary gate vector, and perform dimension expansion on the periodic auxiliary gate vector to obtain a gated tensor.

[0053] The third acquisition module is configured to acquire a previous cell state of a previous time step, and the cell state is used to describe a forgetting degree and a retaining degree of the salinity time sequence.

[0054] The extraction module is configured to extract a trend feature from the previous cell state.

[0055] The third calculation module is configured to obtain a current cell state according to the trend feature, the gated tensor, the input value and the forget value.

[0056] The reconstruction module is configured to activate the current cell state to obtain a current hidden state, and restore the current hidden state to obtain a current salinity data.

[0057] The optimization module is configured to input the current salinity data into a prediction model to obtain a prediction output value, calculate a loss value according to the prediction output value and a training value, adjust the prediction model according to the loss value to obtain a prediction model parameter when the loss function is minimum, adjust the prediction model according to the prediction model parameter to obtain an optimized prediction model.

[0058] The prediction module is configured to input an actual monthly salinity dataset into the optimized prediction model to obtain a predicted salinity value.

[0059] A terminal device includes a memory and a processor, the memory stores a computer program capable of running on the processor, and when the processor loads and executes the computer program, a method for predicting long-time ocean surface salinity is adopted.

[0060] A computer readable storage medium, the computer readable storage medium stores a computer program, the computer program is loaded and executed by a processor, and a method for long time series of ocean surface salinity prediction is adopted.

[0061] The beneficial effects of the present application are:

[0062] By inputting the monthly salinity data into the input gate and the forgetting gate, input values and forgetting values are obtained, and a multi-scale periodic encoding vector is generated according to the time step; the multi-scale periodic encoding vector is input into a fully connected neural network to obtain a periodic auxiliary gate vector, and dimension expansion is performed to obtain a gating tensor; trend features are extracted from the previous cell state; the current cell state is obtained according to the trend features and the gating tensor; the current hidden state is obtained by activating the current cell state and restoring reconstruction to obtain the current salinity data; the current salinity data is input into the prediction model to obtain a prediction output value, and a loss function is calculated according to the prediction output value and the actual value; the prediction model is modulated according to the loss function to obtain the prediction model parameters when the loss function is minimum, and the prediction model is adjusted according to the prediction model parameters to obtain the optimized prediction model; the actual monthly salinity data is input into the optimized prediction model to obtain the predicted salinity value. The present application can improve the accuracy of ocean surface salinity prediction without a large number of parameters. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 It is a structural schematic diagram of the convolution-memory fusion prediction network of the present application;

[0064] Figure 2 It is a prediction result graph and an actual salinity distribution graph of July 2020. DETAILED DESCRIPTION

[0065] A method for long time series of ocean surface salinity prediction, comprising:

[0066] S1, obtaining daily brightness temperature data in a preset time period, and preprocessing the daily brightness temperature data to obtain a daily brightness temperature data set;

[0067] S2, obtaining corresponding monthly salinity data by inverting the daily brightness temperature data set, and different time monthly salinity data forming a monthly salinity data set, wherein the monthly salinity data includes salinity data corresponding to each geographical grid position in a set regional latitude and longitude range within a month;

[0068] Obtaining corresponding monthly salinity data by inverting the daily brightness temperature data set, and different time monthly salinity data forming a monthly salinity data set, wherein the monthly salinity data includes salinity data corresponding to each geographical grid position in a set regional latitude and longitude range within a month;

[0069] The daily brightness temperature data is corrected to obtain corrected data;

[0070] The corrected data is selected and normalized to obtain normalized data;

[0071] According to the normalized data inversion, corresponding monthly salinity data is obtained, and monthly salinity data at different times forms a monthly salinity data set.

[0072] Specifically, daily brightness temperature data observed by the current satellite in recent years is obtained, the data including longitude and latitude corresponding to a grid point in a target region, a sea and land mask, and brightness temperature information, daily brightness temperature data is preprocessed, including data correction, region selection, and normalization operation, to form a brightness temperature data set of the target marine region.

[0073] Using the existing inversion model, the salinity of each month in the past years is inverted into a monthly salinity data set, and the monthly salinity data set is cleaned, transformed, arranged, and normalized to form a three-dimensional monthly salinity data set of the target marine region organized by month. Specifically, a single sample of the sea surface salinity data set is from to monthly salinity data, the size of a single monthly salinity data is X (latitude coordinate index) x Y (longitude coordinate index) x 1 (time step), and each element in the monthly salinity data represents the sea surface salinity value of a specific geographic location and each day in the month;

[0074]

[0075] wherein, is the monthly salinity data set, y is the year, T is the time step, X is the latitude coordinate index, Y is the longitude coordinate index, and R is the real number set.

[0076] S3, combining the monthly salinity data set according to the time step, obtaining a salinity time sequence, and sequentially inputting the salinity time sequence according to the time step into the input gate and the forget gate to obtain the input value and the forgetting value;

[0077] combining the monthly salinity data set according to the time step, obtaining a salinity time sequence, and sequentially inputting the salinity time sequence according to the time step into the input gate and the forget gate to obtain the input value and the forgetting value including:

[0078] setting the time step;

[0079] combining the monthly salinity data set according to the time step, obtaining a salinity time sequence;

[0080] obtaining the hidden state of the previous time step;

[0081] sequentially inputting the salinity time sequence and the hidden state of the previous time step into the input gate and the forget gate according to the time step to obtain the input value and the forgetting value, which is represented as:

[0082]

[0083]

[0084] where, is the input value, is the forget value, is the convolution operation, is the Sigmoid activation function, output range is [0, 1], , , and are the convolution kernels for input and history state pair-gated, is the hidden state of the previous time step, is the salinity time series.

[0085] At each time step, the input gate and the forget gate take the current frame input and the history hidden state as input, and pass through the convolution operation and the sigmoid activation function to generate the gate tensors and . Among them, controls the degree of writing new input information to the current memory state, the threshold is [0, 1], the value is closer to 1, indicating that the current input is more important, the value is closer to 0, indicating that the current input is ignored; determines the retention degree of the memory state of the previous time step, the value is larger, the current history memory continues to pass and retain , the value is smaller, the current time step "forgets" the past content, and discards ; is the convolution operation; , , and are the convolution kernels for input and history state pair-gated, and the convolution kernel size is , , and , here the size of the convolution kernel is k=3; and are the bias terms, is the number of output channels of the gate, and C is the number of hidden channels; is the hidden state, which is obtained by converting the cell state. Through the gating mechanism, the model can dynamically adjust the fusion of new and old information, effectively modeling the time continuity and spatial evolution pattern of the salinity field.

[0086] At each time step, the input and the previous hidden state Together they determine the gating mechanism (input gate) Forgotten Gate Periodic gate and candidate states The calculation is then performed. Next, a gating mechanism controls the inflow and forgetting of information to generate an updated cell state. It serves as the basis for long-term memory at the current time step and is used to generate new hidden states. This provides support for the next time step.

[0087] Specifically, this study innovatively combines the spatiotemporal distribution characteristics of long-term sea surface salinity data to construct a convolutional-memory fusion prediction network. This network structure integrates the mechanisms of Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM). Compared to traditional single-layer ConvLSTM models, this structure, while maintaining spatial structural integrity, closes the output gate and innovatively introduces an additional gating mechanism incorporating multi-scale periodic auxiliary information to handle long-term salinity trends. This mechanism incorporates temporal auxiliary information and embeds it into the gating calculation process, enabling the model to perceive temporal phase and periodic modulation. This results in more accurate and stable predictions of long-term sea surface salinity trends, making it more suitable for high-precision prediction of long-term marine variables. The structure of the convolutional-memory fusion prediction network is as follows: Figure 1 As shown.

[0088] Input layer: Reconstructs the input sea surface salinity dataset to match the network's input format requirements. Specifically, it reconstructs the monthly salinity dataset. It needs to be combined into a single-channel data sequence according to the time series, and the reconstructed input size is... The sea surface salinity sequence is such that its size facilitates the simultaneous extraction of temporal and spatial features, enabling the output layer to generate single-channel salinity predictions for targets at the required time. .

[0089] Multimodal Spatiotemporal Fusion Layer: The multimodal spatiotemporal fusion layer consists of multiple spatiotemporal fusion units (i.e., optimized ConvLSTM units). Each spatiotemporal fusion unit differs from a traditional ConvLSTM unit in that it disables the output gate mechanism and adds a gating mechanism that incorporates multi-scale periodic auxiliary information. Specifically, this layer processes the reconstructed input sea surface salinity sequence as follows:

[0090] 1. Time step input: Input the three-dimensional salinity sequence The data is fed into the spatiotemporal fusion unit frame by frame according to time step t∈[0, T-1]. A two-dimensional salinity sequence is input at each time step. Its size is [1, X, Y], and it also incorporates the hidden state from the previous time step. (C, X, Y) and cell state (C, X, Y) jointly participate in the state update of the current moment.

[0091] 2. Gating mechanism calculation and state update: at each time step, each spatio-temporal fusion unit sequentially completes the following operations:

[0092] ① Calculate the input gate and the forget gate to control the degree of adoption of the current input to new memory and the degree of preservation of the historical state, and the control gate controls whether the current input is adopted, and the forget gate decides whether the memory of the last moment is continued to be retained.

[0093] Specifically, represents the input value, and controls how much the current input is written, represents the forget value, and controls how much the historical memory is retained.

[0094] S4, generating a multi-scale periodic encoding vector according to the time step;

[0095] Generating a multi-scale periodic encoding vector according to the time step includes:

[0096] Obtaining a multi-scale periodic encoding vector formula;

[0097] Inputting the time step into the multi-scale periodic encoding vector formula to obtain the multi-scale periodic encoding vector, and the multi-scale periodic encoding vector formula is represented as:

[0098]

[0099]

[0100] wherein, is a defined period, k is a scale number, k=1, 2 or 3, t represents the difference between the current year and the starting year, is a multi-scale periodic encoding vector, and the starting year is the earliest year in the monthly salinity data set.

[0101] Specifically, when K=1, it represents the annual change period =1, when K=2, it represents the medium-term change , when K=3, it represents the long-term change trend = the annual span of the entire data.

[0102] S5, inputting the multi-scale periodic encoding vector into the fully connected neural network to obtain a period auxiliary gate vector, and dimensionally expanding the period auxiliary gate vector to obtain a gating tensor;

[0103] ​​The multi-scale periodicity coding vector is input into a full connection neural network to obtain a periodic auxiliary gate vector, the periodic auxiliary gate vector is dimensionally expanded to obtain a gating tensor, and the gating tensor comprises:

[0104] The multi-scale periodicity coding vector is input into a two-layer full connection neural network to obtain a periodic auxiliary gate vector, and the periodic auxiliary gate vector is represented as:

[0105]

[0106] wherein, and is a full connection network mapping, and is a bias term matched with the current space state, is a Sigmoid activation function, is used for nonlinear transformation, so that the model can more flexibly process periodicity coding information, is a multi-scale periodicity coding vector.

[0107] The periodic auxiliary gate vector is dimensionally expanded to obtain a gating tensor.

[0108] Specifically, the periodic auxiliary gate vector is calculated, the multi-scale periodicity coding vector is input into a two-layer full connection neural network to obtain a periodic auxiliary gate vector , and the vector can reflect the degree of periodic trend information that each channel should retain.

[0109] S6, obtaining a previous cell state of a previous time step, the cell state is used for describing a forgetting degree and a retaining degree of the salinity time sequence;

[0110] Specifically, an initial hidden state is set as a 0 tensor, and a cell state (the size is [C, X, Y], wherein the initial cell state is set as a 0 tensor.

[0111] S7, extracting a trend feature from the previous cell state;

[0112] The trend feature extracted from the previous cell state is represented as:

[0113]

[0114] wherein, is a convolution kernel used for trend extraction, is the previous cell state, is a hyperbolic tangent activation function, and the output range is [-1, 1], is the trend feature, is a bias term.

[0115] S8, obtaining a current cell state according to the trend feature, the gate tensor, an input value, and a forgetting value;

[0116] Specifically, a periodic gate is fused with a current candidate state to form a new cell state .

[0117] The current cell state obtained according to the trend feature and the gate tensor is represented as:

[0118]

[0119] wherein, is a previous cell state, represents a forgetting value, is a periodic response, is an element-wise multiplication, is an input value.

[0120] Specifically, a periodic auxiliary gate tensor is used The trend feature is adjusted to obtain a periodic response wherein, is an element-wise multiplication (Hadamard Product).

[0121]

[0122] S9, activating the current cell state to obtain a current hidden state, and restoring and reconstructing the current hidden state to obtain current salinity data;

[0123] Specifically, the current cell state is activated to obtain a current hidden state, which is represented as:

[0124]

[0125] wherein, is a current cell state, is a current hidden state, is an activation function

[0126] Specifically, as the time step advances, the network model gradually accumulates the temporal and spatial information, and finally outputs a hidden state sequence of [T, C, X, Y] size, i.e., a dynamic representation after multi-modal fusion. The sequence will serve as the input feature of the subsequent output layer, providing support for the sea surface salinity prediction of the target time step.

[0127] In order to decode the specific salinity value from the learned hidden features, a two-dimensional convolution layer (Conv2D) is used as the output layer. The function of this layer is to decode the hidden features of each time step a hidden state of C channels nonlinearly maps and restores to a single channel representing the current salinity data of the salinity sequence to be predicted The operation is represented as:

[0128]

[0129] wherein, is the current salinity data.

[0130] S10, input the current salinity data into the prediction model to obtain a prediction output value, calculate a loss value according to the prediction output value and a training value, adjust the prediction model according to the loss value to obtain a prediction model parameter at which the loss function is minimized, and adjust the prediction model according to the prediction model parameter to obtain an optimized prediction model;

[0131] The loss value is calculated by a loss function, and the loss function is represented as:

[0132]

[0133] wherein, is the loss value, is a predicted salinity value at a coordinate corresponding position, is a true salinity label value at a coordinate corresponding position, is a value of a mask matrix at a position The land mask is 0 and the ocean mask is 1. In the sea surface salinity prediction task, the data grid after region selection contains both ocean regions and land regions, and the salinity data on land has no physical meaning. If all salinity data is used for loss calculation without distinction, the network model will learn invalid region features, affecting the prediction performance. Using the loss function can make the network learn only the salinity data of the ocean region, thereby improving the training accuracy and model generalization ability.

[0134] The loss function indirectly acts on the cell state of each time step in the model through the backpropagation mechanism, thereby optimizing the entire spatiotemporal memory process and improving the prediction accuracy of future sea surface salinity.

[0135] S11, input the actual monthly salinity data set into the optimized prediction model to obtain a predicted salinity value.

[0136] S11, input the actual monthly salinity data set into the optimized prediction model to obtain a predicted salinity value.

[0137] ​Specifically, in the present embodiment, the brightness temperature data of SMOS satellite in June, July and August during 2015-2010 is obtained (including longitude and latitude grid information, land mask and brightness temperature observation value) and a continuous daily brightness temperature data set is formed.

[0138] The daily brightness temperature data is preprocessed, specifically, the data is corrected, abnormal or distorted data is removed; the region is selected, the brightness temperature data of the target sea area range is extracted; the normalization processing is performed, and the brightness temperature value is mapped to a unified scale to facilitate subsequent model training.

[0139] A continuous daily brightness temperature data set of the target area is formed, and the data structure is a three-dimensional array of [longitude, latitude, time]. The existing sea surface salinity inversion model (based on daily brightness temperature data) is used to monthly invert the corresponding sea surface salinity data in June, July and August from 2015 to 2020, and a monthly salinity data set for June, July and August in each year from 2015 to 2020 is formed.

[0140] The salinity data is normalized according to time and divided into training set, validation set and test set, 80% of the monthly salinity data in June and August from 2015 to 2020 is used as the training set; 20% of the monthly salinity data in June and August from 2015 to 2019 is used as the validation set; the monthly salinity data in July from 2015 to 2020 is used as the test set, and the data in July from 2015 to 2019 is used to predict the data in July 2020.

[0141] A training convolution-memory fusion prediction network is constructed, the monthly salinity data is reconstructed into a sequence input format of [T,1,X,Y], the monthly salinity sequence of the training set is input, and the network is fed step by step to learn features, using MSE (mean square error) as the loss function, calculating the error between the predicted salinity and the true salinity label, updating the network parameters through the back propagation algorithm, and continuously training until the loss function converges or reaches the set training number of rounds.

[0142] The trained convolution-memory fusion prediction model is used to input the salinity data sequence that has not been trained on the network structure, i.e. the data in July from 2015 to 2019, to predict the sea surface salinity distribution in July 2020. The prediction results are saved after being denormalized, and compared with the real salinity observation data in July 2020, the mean square error (MSE) between the predicted value and the actual value is calculated, and the comparison distribution graph of the predicted value and the actual value of the sea surface salinity in July 2020 is drawn.

[0143] Figure 2The prediction result graph and the actual salinity distribution graph in July 2020 are given. Through the group of graphs, it can be found that although there is a certain difference between the predicted value and the actual value, the overall change trend is consistent, indicating that the prediction model has good accuracy in capturing the salinity distribution trend, can extract the most important features from the data, and make accurate prediction based on these features.

[0144] A long-time marine surface salinity prediction system comprises.

[0145] The first acquisition module is configured to acquire diurnal skin temperature data in a preset time period, and preprocess the diurnal skin temperature data to obtain a diurnal skin temperature data set.

[0146] The second acquisition module is configured to obtain corresponding monthly salinity data by inverting the diurnal skin temperature data set, and different time monthly salinity data constitutes a monthly salinity data set, wherein the monthly salinity data includes salinity data corresponding to each geographical grid position in a set regional latitude and longitude range within a month.

[0147] The first calculation module is configured to combine the monthly salinity data set according to time steps to obtain a salinity time sequence, and input the salinity time sequence into an input gate and a forget gate in sequence according to time steps to obtain an input value and a forget value.

[0148] The second calculation module is configured to generate a multi-scale periodic encoding vector according to the time steps.

[0149] The gated tensor generation module is configured to input the multi-scale periodic encoding vector into a fully connected neural network to obtain a periodic auxiliary gate vector, and perform dimension expansion on the periodic auxiliary gate vector to obtain a gated tensor.

[0150] The third acquisition module is configured to acquire a previous cell state of a previous time step, and the cell state is used to describe the forgetting degree and the retaining degree of the salinity time sequence.

[0151] The extraction module is configured to extract a trend feature from the previous cell state.

[0152] The third calculation module is configured to obtain a current cell state according to the trend feature, the gated tensor, the input value and the forget value.

[0153] The reconstruction module is configured to activate the current cell state to obtain a current hidden state, and restore the current hidden state to obtain a current salinity data.

[0154] an optimization module, configured to input the current salinity data into a prediction model to obtain a predicted output value, calculate a loss value according to the predicted output value and a training value, adjust the prediction model according to the loss value to obtain a prediction model parameter at which the loss function is minimized, and adjust the prediction model according to the prediction model parameter to obtain an optimized prediction model;

[0155] a prediction module, configured to input an actual monthly salinity data set into the optimized prediction model to obtain a predicted salinity value.

[0156] The embodiment of the present application also discloses a terminal device, including a memory and a processor, the memory stores a computer program capable of running on the processor, and when the processor loads and executes the computer program, a method for predicting long-time ocean surface salinity is adopted.

[0157] The terminal device can be a computer device such as a desktop computer, a notebook computer or a cloud server, and the terminal device includes but is not limited to a processor and a memory. For example, the terminal device can also include an input / output device, a network access device and a bus.

[0158] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), ready programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, and the present application does not limit it.

[0159] The memory can be an internal storage unit of the terminal device, such as a hard disk or a memory of the terminal device, or an external storage device of the terminal device, such as a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD) or a flash memory card (FC) equipped on the terminal device. The memory can also be a combination of the internal storage unit and the external storage device of the terminal device. The memory is used to store computer programs and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or will be output, and the present application does not limit it.

[0160] The terminal device stores the method for predicting long-time ocean surface salinity in the memory of the terminal device, and loads and executes it on the processor of the terminal device, which is convenient to use.

[0161] The embodiment of the application further discloses a computer readable storage medium, and the computer readable storage medium stores a computer program.

[0162] The computer program can be stored in the computer readable medium, and the computer program includes computer program code. The computer program code can be in the form of source code, object code, executable files or some intermediate forms, etc. The computer readable medium includes any entity or device, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier wave signal, telecommunication signal and software distribution medium, etc. that can carry the computer program code. It should be noted that the computer readable medium includes but is not limited to the above components.

[0163] The computer readable storage medium stores the method for predicting long-time ocean surface salinity in one of the above embodiments, and is loaded and executed on the processor to facilitate storage and application of the method.

[0164] It should be understood by those skilled in the art that the above discussion of any embodiment is only exemplary and is not intended to imply that the protection scope of the application is limited to these examples; under the idea of the application, the above embodiments or technical features in different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of one or more embodiments of the application as described above. In order to be brief, they are not provided in details.

[0165] One or more embodiments of the application are intended to cover all such alternatives, modifications and variations falling within the broad scope of the application. Therefore, any omission, modification, equivalent replacement, improvement, etc. made in the spirit and principles of one or more embodiments of the application should be included in the protection scope of the application.

Claims

1. A method for predicting ocean surface salinity over long time series, characterized in that, include: Obtain the solar brightness temperature data within a preset time period, and preprocess the solar brightness temperature data to obtain a solar brightness temperature dataset; The corresponding monthly salinity data is obtained by inverting the solar brightness temperature dataset. The monthly salinity data at different times constitutes the monthly salinity dataset. The monthly salinity data includes the salinity data corresponding to each geographic grid location within a set latitude and longitude range within a month. The monthly salinity dataset is combined according to time steps to obtain a salinity time series. The salinity time series is then input into the input gate and the forget gate sequentially according to time steps to obtain the input value and the forgotten value. Based on the time step, a multi-scale periodic encoding vector is generated; The multi-scale periodic encoding vector is input into a fully connected neural network to obtain a periodic auxiliary gate vector. The periodic auxiliary gate vector is then expanded in dimension to obtain a gated tensor. Obtain the previous cell state from the previous time step. The cell state is used to describe the degree of forgetting and retention of the salinity time series. Extract trend features from the pre-cellular state; The current cell state is obtained based on the trend characteristics, the gating tensor, the input value, and the forgotten value. The current cell state is activated to obtain the current hidden state, and the current hidden state is restored and reconstructed to obtain the current salinity data; The current salinity data is input into the prediction model to obtain the prediction output value. Based on the prediction output value and the training value, the loss value is calculated. The prediction model is adjusted based on the loss value to obtain the prediction model parameters when the loss value is minimized. The prediction model is then adjusted based on the prediction model parameters to obtain the optimized prediction model. The actual monthly salinity dataset is input into the optimized prediction model to obtain the predicted salinity value.

2. The method for predicting ocean surface salinity over long time periods as described in claim 1, characterized in that, The monthly salinity data is obtained by inverting the solar brightness temperature dataset. The monthly salinity dataset, composed of monthly salinity data from different times, includes: The solar brightness temperature data is corrected to obtain the corrected data; The corrected data is subjected to region selection and normalization operations to obtain normalized data; The corresponding monthly salinity data is obtained by inverting the normalized data, and the monthly salinity data at different times constitute the monthly salinity dataset.

3. The method for predicting ocean surface salinity over long time periods as described in claim 1, characterized in that, The step of combining the monthly salinity dataset according to time steps to obtain a salinity time series, and then sequentially inputting the salinity time series into an input gate and a forgetting gate according to time steps to obtain input values ​​and forgotten values ​​includes: Set time steps; The monthly salinity dataset is combined according to time steps to obtain a salinity time series; Get the hidden state of the previous time step; The salinity time series and the hidden state of the previous time step are sequentially input into the input gate and the forget gate according to the time step to obtain the input value and the forgotten value, represented as: in, For input values, Indicates the forgotten value. This represents the convolution operation. This represents the Sigmoid activation function, with an output range of [0,1]. , , and These are the convolution kernels that are gated by the input and the historical state, respectively. This is the hidden state from the previous time step. and For bias terms, This is a time series of salinity.

4. The method for predicting ocean surface salinity over long time periods as described in claim 1, characterized in that, The step of generating a multi-scale periodic encoding vector based on the time step includes: Formula for obtaining multi-scale periodic encoding vectors; The time step is input into the multi-scale periodic coding vector formula to obtain the multi-scale periodic coding vector, which is expressed as follows: in, For the defined period, k is the scale number, k=1, 2 or 3, and t represents the difference between the current year and the starting year. It is a multi-scale periodic encoding vector, with the starting year being the earliest year in the monthly salinity dataset.

5. The method for predicting ocean surface salinity over long time periods as described in claim 1, characterized in that, The step of inputting the multi-scale periodic encoding vector into a fully connected neural network to obtain a periodic auxiliary gate vector, and then expanding the dimension of the periodic auxiliary gate vector to obtain a gated tensor includes: The multi-scale periodic encoding vector is input into a two-layer fully connected neural network to obtain a periodic auxiliary gate vector, which is expressed as: in, and For fully connected network mapping, and For the bias term that matches the current spatial state, It is the Sigmoid activation function. It is used for nonlinear transformations, enabling the model to handle periodic encoded information more flexibly. It is a multi-scale periodic encoding vector; The periodic auxiliary gate vector is expanded in dimension to obtain the gate tensor.

6. The method for predicting ocean surface salinity over long time periods as described in claim 1, characterized in that, The trend features extracted from the pre-cellular state are represented as follows: in, For trend extraction, the convolution kernel, It is in the pre-cellular state. This is a hyperbolic tangent activation function with an output range of [-1, 1]. As a trend feature, This is a bias term.

7. The method for predicting ocean surface salinity over long time periods as described in claim 1, characterized in that, Based on the trend characteristics and the gating tensor, the current cell state is represented as follows: in, It is in the pre-cellular state. Indicates the forgotten value. It is a periodic response. For element-wise multiplication, For input values.

8. A system for predicting ocean surface salinity over long time series, characterized in that, include: The first acquisition module is used to acquire solar brightness temperature data within a preset time period and preprocess the solar brightness temperature data to obtain a solar brightness temperature dataset. The second acquisition module is used to invert the solar brightness temperature dataset to obtain the corresponding monthly salinity data. The monthly salinity data at different times constitutes the monthly salinity dataset. The monthly salinity data includes the salinity data corresponding to each geographical grid location within a set latitude and longitude range within a month. The first calculation module is used to combine the monthly salinity dataset according to time steps to obtain a salinity time series, and to input the salinity time series into the input gate and the forget gate in sequence according to time steps to obtain the input value and the forgotten value; The second calculation module is used to generate a multi-scale periodic encoding vector based on the time step. The gated tensor generation module is used to input the multi-scale periodic encoded vector into a fully connected neural network to obtain a periodic auxiliary gate vector, and to expand the dimension of the periodic auxiliary gate vector to obtain a gated tensor. The third acquisition module is used to acquire the previous cell state at the previous time step. The cell state is used to describe the degree of forgetting and retention of the salinity time series. An extraction module is used to extract trend features from the pre-cell state; The third calculation module is used to obtain the current cell state based on the trend features, the gating tensor, the input value, and the forgotten value. The reconstruction module is used to activate the current cell state to obtain the current hidden state, and restore and reconstruct the current hidden state to obtain the current salinity data; The optimization module is used to input the current salinity data into the prediction model to obtain the prediction output value, calculate the loss value based on the prediction output value and the training value, adjust the prediction model based on the loss value to obtain the prediction model parameters when the loss value is minimized, and adjust the prediction model based on the prediction model parameters to obtain the optimized prediction model. The prediction module is used to input the actual monthly salinity dataset into the optimized prediction model to obtain the predicted salinity value.

9. A terminal device, comprising a memory and a processor, characterized in that, The memory stores a computer program that can run on a processor, and when the processor loads and executes the computer program, it employs the method described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is loaded and executed by the processor, it employs the method described in any one of claims 1 to 7.

Images