A hybrid model sea ice extent prediction method based on feature-focused capture mechanism
Through a hybrid model with feature-focused capture mechanism, combined with CNN, BiTCN and BiGRU, the problems of single sea ice extent prediction model and low accuracy in melting season are solved, high-precision prediction of sea ice extent is achieved, and the modeling ability of complex nonlinear relationships and time series dependencies is enhanced.
Patent Information
- Application Number
- CN202411492590.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-24
- Publication Date
- 2025-09-30
- Estimated Expiration
- 2044-10-24
AI Technical Summary
Existing sea ice extent prediction methods have a single model, low accuracy in sea ice prediction during the melt season, lack of complex nonlinear relationship extraction, and ignore differences in the importance of features.
A hybrid model based on the feature focus capture mechanism is adopted, including a spatial feature extraction module, a bidirectional local feature extraction module and a bidirectional temporal feature extraction module. Combined with the attention mechanism, the convolutional neural network (CNN), the bidirectional temporal convolutional network (BiTCN) and the bidirectional gated recurrent unit (BiGRU) are used to extract and fuse the features of sea ice concentration and atmospheric and meteorological variables, and dynamically adjust the feature weights.
The overall accuracy of sea ice extent prediction has been improved, especially during the melt season, and the ability to perceive changes in sea ice extent has been enhanced, dynamically adapting to changes in feature importance and reducing interference from irrelevant information.
Smart Images

Figure CN119761540B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of satellite remote sensing and deep learning technology, and specifically to a hybrid model sea ice extent prediction method based on a feature-focused capture mechanism. Background Art
[0002] Changes in Arctic sea ice extent are closely related to the stability of marine ecosystems, the global carbon cycle, the survival of Arctic wildlife, and the production and life of human society. Therefore, predicting Arctic sea ice extent is of great significance.
[0003] Sea ice extent prediction methods include those based on statistical data, dynamical numerical models, and machine learning. Statistical methods use statistical methods to predict Arctic SIE, finding that sea ice concentration (SIC) is a key predictor of SIE. Statistical methods primarily use relationships between historical observational data (such as extent and density), atmospheric conditions (such as temperature, sea level pressure, and clouds), and oceanic conditions (such as sea surface temperature) to establish statistical models of temporal or characteristic changes in sea ice variables (such as concentration, extent, ice type, and thickness) to predict future sea ice extent. Their advantage lies in their computational simplicity, but their disadvantage lies in their lack of consideration of the physical mechanisms of sea ice evolution. Dynamical numerical models rely on an understanding of the physics of the climate system. Within the constraints of the model's dynamical framework, initial and boundary conditions, they utilize complex numerical methods to solve the equations of motion, thermodynamics, and radiation to simulate sea ice changes. Their advantage lies in their high interpretability of predictions. However, their disadvantages lie in the incompleteness of parameterization schemes for the processes and mechanisms of sea ice change, the large uncertainty in predictions, the high computational complexity, and their strong dependence on the initial sea ice state.
[0004] Thanks to the release of high-quality satellite remote sensing data and atmospheric-ocean reanalysis data from the National Snow and Ice Data Center (NSIDC), machine learning-based sea ice extent prediction methods have become a popular research topic in recent years. These methods are primarily categorized into univariate and multivariate machine learning-based methods. Univariate methods use only past sea ice extent data as a variable to predict sea ice extent. As sea ice research has advanced, numerous researchers have observed and confirmed the complex relationship between sea ice and atmospheric and meteorological variables (such as wind speed, sea level pressure, temperature, and long- and shortwave radiation). Based on this research, multivariate machine learning-based sea ice extent prediction methods, which integrate atmospheric and meteorological variables, have been increasingly studied. However, existing machine learning methods are single-model methods and fail to fully consider the important factor of sea ice concentration. Furthermore, existing algorithms perform poorly during the sea ice melt season (May to October each year). In particular, improving the accuracy of SIE predictions during September, the month of minimum sea ice extent, remains a challenge for current algorithms. In view of this, the present invention proposes a hybrid model sea ice extent prediction method based on feature-focused capture mechanism. Summary of the Invention
[0005] The purpose of the present invention is to propose a hybrid model sea ice extent prediction method based on a feature focus capture mechanism to solve the problems of existing sea ice extent prediction methods, such as a single model, low sea ice prediction accuracy in the melt season, lack of complex nonlinear relationship extraction, and ignoring the importance differences between features.
[0006] In order to achieve the above object, the present invention adopts the following technical solutions:
[0007] A hybrid model sea ice extent prediction method based on a feature-focused capture mechanism includes the following steps:
[0008] S1. Obtain monthly EAR5 reanalysis data of sea ice concentration and atmospheric and meteorological variables as experimental data, obtain monthly sea ice extent data from the National Snow and Ice Data Center (NSIDC) as reference data, and preprocess the experimental and reference data to obtain the required standard data;
[0009] S2. Use the spatial feature extraction module to extract the spatial features hidden in the two-dimensional sea ice concentration data, and fuse the extracted features with subsequent features as important predictors to predict future sea ice extent;
[0010] S3: Fuse the spatial features extracted in S2 with the atmospheric and meteorological multivariate features, and input the fused tensor into the bidirectional local feature extraction module to capture the bidirectional relationships and dependencies that only exist in a specific time or region, such as the periodicity, trend, and peak of the sequence changes;
[0011] S4, using the bidirectional temporal feature extraction module to continue processing the features extracted by the bidirectional local feature extraction module in S3, to capture the complex relationship between the moments before and after the features;
[0012] S5. Design a feature-focused capture mechanism to give higher weight to relevant information and reduce the interference of irrelevant information.
[0013] Preferably, the monthly sea ice extent data obtained from the National Snow and Ice Data Center (NSIDC) and the atmospheric, meteorological, and sea ice concentration data obtained from EAR5 may have some missing values. For the experimental data and reference data obtained in S1, cubic spline interpolation is used to replace the missing values. The cubic spline interpolation is defined as follows:
[0014] For the known n+1 data points (x0, y0)(x1, y1),…,(x n ,y n ), the cubic polynomial S between each adjacent data point i (x) is expressed as: S i (x) = a i +b i (xx i )+c i (xx i ) 2 +d i (xx i ) 3 , and meet the following conditions:
[0015] (1)S i (x) is a cubic polynomial in each small interval;
[0016] (2)S i (x) has the same first-order and second-order derivatives at the connection between two adjacent small intervals, that is,
[0017] ″″″
[0018] S i (x i )=S i (x i+1 ),S i (x i )=S i (x i+1 );
[0019] (3)S i (x) passes through all given data points, i.e. S i (x i )=y i ;
[0020] ″″
[0021] (4) The boundary conditions are S0(x0)=0 and S n (x n )=0;
[0022] Based on the above conditions, establish a linear equation system to solve the coefficient a i ,b i ,c i ,d i ;
[0023] After solving, when interpolating at a given point x, determine the interval [x i ,x i+1 ], using the corresponding cubic polynomial S i (x) is calculated to obtain the corresponding interpolation value.
[0024] Preferably, since different data have different dimensions, if data standardization is not performed, some characteristic data may be ignored. The experimental data and reference data are preprocessed as described in S1 to obtain the required standard data, specifically including the following:
[0025] The data is standardized using the MinMax standardization technique. The MinMax standardization formula is as follows:
[0026]
[0027] Among them, Data represents the original data; Data min Indicates the minimum value in the original data; Data max Indicates the maximum value in the original data;
[0028] After the model predicts the sea ice extent data, the data is denormalized for observation; the denormalization formula is as follows:
[0029]
[0030] Among them, Data * represents the denormalized value, Represents the model prediction value.
[0031] Preferably, the spatial feature extraction module in S2 is implemented by a convolutional neural network (CNN), and the specific design is as follows:
[0032] The convolutional neural network (CNN) of the spatial feature extraction module includes a four-convolutional layer block structure. Each convolutional layer block includes a convolutional layer, a weight normalization layer, a ReLU layer, and a maximum pooling layer. The convolution kernel size of the convolution layer is 3×3. The convolution layer identifies spatial features such as edges and textures in the sea ice concentration data through convolution operations. The maximum pooling layer performs a pooling operation with a 2×2 window with a stride of 2 to reduce the dimension of the features while retaining the main features.
[0033] The data format of the input convolutional neural network (CNN) is (10, 1, 1440, 261), where 10 is the batch size, 1 is the number of channels, and 1440 × 261 represents the latitude and longitude of the data. Multi-level feature extraction and dimensionality reduction are performed on the data, and the output feature shape is (10, 64, 90, 16).
[0034] The obtained feature representation is passed to two fully connected layers for further processing to output spatial features, and the final output feature shape is (10,64); the spatial features extracted by convolutional neural network (CNN) can more accurately capture the distribution pattern and spatial change trend of sea ice. These features can be fused with subsequent features as important predictors to predict future sea ice extent.
[0035] Preferably, the bidirectional local feature extraction module in S3 is implemented by a bidirectional temporal convolutional network (BiTCN), and the specific design is as follows:
[0036] Bidirectional temporal convolutional network (BiTCN) consists of two components: forward temporal convolution and reverse temporal convolution. Forward temporal convolution and reverse temporal convolution extract features of different scales by stacking multiple convolutional layers, thereby capturing features of different time ranges in sequence data.
[0037] The above-mentioned multi-scale information extraction capability makes TCN more sensitive to local dependencies in sequence data and effectively extracts local features of time series;
[0038] The TCN consists of three parts: causal convolution, dilated convolution, and residual connection. Causal convolution and dilated convolution are combined to form causal dilated convolution. The causal dilated convolution is defined as:
[0039] Filter F=(f1,f2,…,f K ), time series X=(x1,x2,…,x T ), in x tThe dilated convolution formula with a dilation rate of d is as follows:
[0040]
[0041] The residual network is used to solve the problems of gradient explosion and gradient disappearance caused by network deepening. It directly adds the input of each layer to the output. The output formula is as follows:
[0042] o=ReLU(X+F(X)) (4)
[0043] Where o represents the output of the residual module; X represents the input data; F(X) represents the data after nonlinear transformation;
[0044] The causal dilation convolution has a total of 4 layers, the kernel is set to 2, and the dilation coefficient d of each layer is set to 1, 2, 4, and 8 respectively; deeper model layers are used to obtain a higher and larger field of view; the bidirectional temporal convolutional network (BiTCN)'s bidirectional structure allows the model to extract data features from two time directions, more accurately capturing the local changes and dependencies of sequence information.
[0045] Preferably, the bidirectional temporal feature extraction module in S4 is implemented by a bidirectional gated recurrent unit (BiGRU), and the specific design is as follows:
[0046] The Bidirectional Gated Recurrent Unit (BiGRU) consists of two unidirectional GRUs running in opposite directions. The GRU network is trained with two gating units: the update gate and the reset gate.
[0047] The function of the update gate is expressed as follows:
[0048] z t =σ(W z ·[h t-1 ,x t ]+b z ) (5)
[0049] The function of the reset gate is expressed as follows:
[0050] r t =σ(W r ·[h t-1 ,x t ]+b r ) (6)
[0051] The new candidate state formula is as follows:
[0052]
[0053] The formula for updating the hidden state is as follows:
[0054]
[0055] Among them, x t Indicates input data; h t Represents the output of the GRU unit; r t Represents the reset gate; z t represents the update gate; σ represents the Sigmoid function; W z ,W r ,W h Represent the update gate, reset gate and weight matrix of candidate hidden states respectively; select the bidirectional gated recurrent unit (BiGRU), which mines the input features in both sequential and reverse order in the time dimension to extract the complex dependency relationship between the features of the previous and next moments.
[0056] Preferably, the basic idea of the attention mechanism of the feature focus capture mechanism is to filter out useful information through certain means and find factors that are more important to the prediction target, thereby improving the efficiency and reliability of the prediction; the feature focus capture mechanism described in S5 is implemented through the attention mechanism, and the specific design is as follows:
[0057] The weight calculation formula of the feature focus capture mechanism is as follows:
[0058]
[0059] Among them, W t represents the weight matrix; b t represents the bias coefficient; h t Represents the output of the BiGRU layer; l t represents the output of the attention layer;
[0060] The feature-focused capture mechanism is used to enable the model to automatically learn key information in the input sequence. By giving higher weights to relevant information, the model pays more attention to the parts related to the sea ice extent and reduces interference with irrelevant information, thereby more accurately capturing the important information at each time step.
[0061] Compared with the existing technology, the present invention provides a hybrid model sea ice extent prediction method based on the feature-focused capture mechanism, which has the following beneficial effects:
[0062] (1) To overcome the limitations of a single model algorithm structure that restricts prediction accuracy, this paper proposes a hybrid model prediction mechanism. This mechanism is implemented through the following modules: a spatial feature extraction module based on a convolutional neural network (CNN); a bidirectional feature extraction module based on a bidirectional temporal convolutional network (BiTCN) and a bidirectional gated recurrent unit (BiGRU); and a feature focus capture mechanism based on an attention mechanism. The hybrid model leverages each model's strengths and works collaboratively to achieve an overall improvement in the accuracy of sea ice extent prediction.
[0063] (2) In order to solve the problem of low prediction accuracy of sea ice extent in the melt season, the present invention proposes for the first time to explore the relationship between sea ice concentration and sea ice extent. The spatial features of sea ice concentration data are extracted through a spatial feature extraction module and integrated with atmospheric and meteorological multivariate features. Monthly sea ice concentration data provides information on the spatial distribution and changes of sea ice. Feature extraction can capture the monthly spatial correlation of sea ice, thereby enhancing the prediction model's ability to perceive changes in sea ice extent. At the same time, atmospheric and meteorological variables provide environmental background information on the formation and changes of sea ice. Feature fusion can comprehensively reflect the dynamic changes of sea ice, thereby achieving high-precision prediction of sea ice extent in the melt season.
[0064] (3) In order to solve the problem of limited feature extraction capability of unidirectional networks and lack of extraction of complex nonlinear relationships between features, the present invention proposes a bidirectional feature extraction module, including bidirectional local feature extraction and bidirectional temporal feature extraction, which fully mines the local relationships and temporal dependencies between data in both sequential and reverse directions, thereby enhancing the modeling capability of complex nonlinear relationships and temporal dependencies in data.
[0065] (4) In order to overcome the problems of existing algorithms ignoring the importance differences between features and being unable to dynamically adjust feature weights, the present invention proposes a feature focus capture mechanism, which automatically learns the key information in the time series and gives higher weights to related features, so that the model can dynamically adapt to the changes in feature importance of different data and reduce the interference of irrelevant information. BRIEF DESCRIPTION OF THE DRAWINGS
[0066] Figure 1 This is an overall structural diagram of the hybrid model sea ice extent prediction method based on the feature-focused capture mechanism mentioned in Example 1 of the present invention;
[0067] Figure 2 This is a heat map of the Pearson correlation coefficients between the atmospheric and meteorological variables and the sea ice concentration and sea ice extent mentioned in Example 1 of the present invention. DETAILED DESCRIPTION
[0068] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.
[0069] Example 1:
[0070] See also Figure 1 ,This paper proposes a hybrid model sea ice extent prediction algorithm based on ,feature key capture mechanism, such as Figure 1 As shown, the following steps are included:
[0071] Step 1: Obtain monthly ERA5 reanalysis data of sea ice concentration and nine atmospheric and meteorological variables as experimental data, and monthly sea ice extent data from NSIDC as reference data. All data are preprocessed to obtain the required standard data.
[0072] The present invention uses monthly sea ice concentration data and atmospheric and meteorological data covering the Arctic Ocean and adjacent land areas from 1979 to 2021 as experimental data, and monthly sea ice extent data as reference data. The monthly sea ice concentration data comes from the monthly ERA5 reanalysis data of the European Centre for Medium-Range Weather Forecasts (ECMWF). This data set integrates the sea ice concentration monitoring results provided by the Ocean and Sea Ice Satellite Application Facility (OSI SAF) of the European Organization for the Exploitation of Meteorological Satellites (EUMETSAT), including detailed information on the sea ice edge and concentration. The data set is continuously verified by comparing navigation ice maps to ensure the accuracy and reliability of the data. The monthly sea ice concentration data used is in NetCDF format, which organizes data in conventional latitude and longitude grids and optimizes the grid geometry using bilinear interpolation technology. This data grid covers the entire world, with a latitude range of -90 to +90 degrees (with reference to the equator) and a longitude range of 0 to 360 degrees (with reference to the Greenwich prime meridian). The resolution of each grid cell reaches 0.25 degrees.
[0073] Numerous studies have demonstrated the causal relationship between the interaction between Arctic sea ice and atmospheric and meteorological variables such as air pressure, wind speed, and sea surface temperature. Therefore, this paper selected nine atmospheric and meteorological data from the European Centre for Medium-Range Weather Forecasts (ECMWF) monthly ERA5 reanalysis data. These variables include: 10m wind speed (10m U-component of wind, 10m V-component of wind, abbreviated as 10m_wind); surface pressure of the atmosphere over land, ocean, and inland water (Surface pressure); air temperature 2m above the surface of land, ocean, or inland water (2m Temperature, abbreviated as t2m); the amount of thermal (also called longwave or ground) radiation emitted by the atmosphere and clouds reaching the horizontal plane of the Earth's surface (Surface thermal radiation downwards, abbreviated as LW_down); the mean total precipitation rate (rainfall); the mean snowfall rate (snowfall); sea surface temperature (SST); the amount of solar radiation reaching the horizontal plane of the Earth's surface, also called shortwave radiation (Surface solar radiation downwards, abbreviated as SW_down); and the mass of water vapor per kilogram of moist air (Specific humidity).
[0074] Monthly sea ice extent data are derived from the monthly averages of Nimbus-7 SSMR and DMSP SSM / I-SSMIS Passive Microwave Data V2 from the National Snow and Ice Data Center (NSIDC). These data are generated using a NASA Team algorithm developed by the Cryospheric Sciences Laboratory at NASA's Goddard Space Flight Center (GSFC). Monthly sea ice extent data are divided into freezing season data and melt season data. Generally speaking, March is the freezing season, when Arctic sea ice reaches its maximum extent of the year; September is the melt season, when Arctic sea ice reaches its minimum extent of the year.
[0075] Figure 2A heatmap of Pearson correlation coefficients between selected atmospheric and meteorological variables, sea ice concentration, and sea ice extent is shown. The figure shows that all nine atmospheric and meteorological variables, as well as sea ice concentration, have positive or negative correlations with sea ice extent, but their impact on sea ice extent varies. Darker colors and larger absolute values indicate a stronger correlation, while lighter colors and smaller absolute values indicate a weaker correlation. For example, sea surface temperature (SST) exhibits the strongest negative correlation with sea ice extent, while sea ice concentration (SIC) exhibits the strongest positive correlation.
[0076] The data set preparation process in this embodiment is as follows:
[0077] (1) Replace missing values.
[0078] The monthly sea ice extent data obtained from NSIDC and the atmospheric, meteorological, and sea ice concentration data obtained from EAR5 may have some missing values. The reason for the missing values is that some data points are not observed or recorded and the missing values cannot be defined. Therefore, for the integrity of the data, interpolation is used to handle the missing values. Interpolation is a method of inferring the values of unknown locations by using the values of known data points. It can reasonably compensate for the missing values in the data. This paper uses cubic spline interpolation, which uses cubic polynomials between each adjacent data point and ensures that these polynomials have continuous function values and first-order and second-order derivative values at each data point. The basic idea is to approximate the data in a small interval with a low-order polynomial, and use the connection conditions of the junctions of these polynomials to ensure the smoothness of the entire curve. The cubic spline interpolation is defined as follows:
[0079] For the known n+1 data points (x0,y0),(x1,y1),…,(x n ,y n ), the cubic polynomial S between each adjacent data point i (x) is expressed as S i (x) = a i +b i (xx i )+c i (xx i ) 2 +d i (xx i ) 3 , and meet the following conditions: S i (x) is a cubic polynomial in each small interval; S i (x) has the same first-order and second-order derivatives at the connection between two adjacent small intervals, that is, S i '(x i )=S i '(x i+1 ),Si ”(x i )=S i ”(x i+1 );S i (x) passes through all given data points, i.e. S i (x i )=y i The boundary conditions are S″0(x0)=0 and S″ n (x n )=0. Based on these conditions, a linear equation system can be established to solve the coefficient a i ,b i ,c i ,d i .
[0080] After solving, to interpolate at a given point x, you only need to determine the interval [x i ,x i+1 ], using the corresponding cubic polynomial S i (x) is calculated to obtain the corresponding interpolation value.
[0081] (2) Standardization
[0082] Because different data have different dimensions, some features may be overlooked if data is not normalized. Therefore, this article uses Min-Max normalization to normalize the data. Min-Max normalization is a common method for data normalization. It performs a linear transformation on the original data, mapping the original data value into a value in the interval [0, 1]. The Min-Max normalization formula is shown in Equation 1.
[0083]
[0084] Among them, Data represents the original data, Data min Represents the minimum value in the original data, Data max Represents the maximum value in the original data.
[0085] After the model predicts the sea ice extent data, the data needs to be denormalized for observation.
[0086] The denormalization formula is shown in Equation 2.
[0087]
[0088] Among them, Data * represents the denormalized value, Represents the model prediction value.
[0089] Step 2: Use the spatial feature extraction module to extract the spatial features hidden in the two-dimensional sea ice concentration data. This feature will be fused with subsequent features as an important predictor to predict future sea ice extent.
[0090] The spatial feature extraction module used in the present invention is used to extract the hidden spatial features in the two-dimensional sea ice concentration data. The module is implemented by a convolutional neural network (CNN). The spatial features extracted by this module will serve as an important factor in the subsequent sea ice range prediction. The CNN network used in this article contains a total of four convolutional layer blocks. The convolution kernel size of the convolution layer is 3×3, and the convolution layer identifies spatial features such as edges and textures in the sea ice concentration data through convolution operations. The maximum pooling layer performs a pooling operation with a 2×2 window with a step size of 2 to reduce the dimension of the features and retain the main features. The data format input to the CNN is (10, 1, 1440, 261), where 10 is the batch size, 1 is the number of channels, and 1440×261 represents the longitude and latitude of the data. After multi-level feature extraction and dimensionality reduction processing, the output feature shape is (10, 64, 90, 16). Finally, the obtained feature representation is passed to two fully connected layers for further processing to output spatial features, and the final output feature shape is (10, 64).
[0091] The spatial features extracted by CNN can more accurately capture the distribution pattern and spatial variation trend of sea ice. These features can be integrated with subsequent features and used as important predictors to predict future sea ice extent.
[0092] Step 3: The sea ice spatial features extracted from the spatial feature extraction module will be fused with nine atmospheric and meteorological multivariate features. The fused tensor will then be input into the bidirectional local feature extraction module to capture certain bidirectional relationships and dependencies that only exist in a specific time or region, such as the periodicity, trend, and peak of the sequence changes.
[0093] The sea ice spatial features extracted from the spatial feature extraction module are fused with nine atmospheric and meteorological multivariate features. Since the spatial features are in the format of (10, 64), while the original atmospheric and meteorological variables are in the format of (10, 9), to facilitate the fusion of these features, the atmospheric and meteorological variables are mapped to (10, 64) and then concatenated according to the feature dimension, resulting in a fused tensor with a shape of (10, 128). This tensor is then input into the bidirectional local feature extraction module to capture certain bidirectional relationships and dependencies that may only exist in a specific time or region, such as periodicity, trends, and peaks in sequence changes. The bidirectional local feature extraction module is implemented using the BiTCN network. The BiTCN consists of two main components: forward temporal convolution and backward temporal convolution. Temporal convolution extracts features at different scales by stacking multiple convolutional layers, thereby capturing features from different time ranges within the sequence data. This multi-scale information extraction capability makes the TCN more sensitive to local dependencies in the sequence data and can effectively extract local features of the time series. TCN is a process of dilated causal convolution, which consists of three parts: causal convolution, dilated convolution, and residual connection.
[0094] TCN combines causal convolution with dilated convolution and uses causal dilated convolution. Dilated causal convolution is defined as: filter F = (f1,f2,····,f K ), time series X=(x1,x2,····,x T ), in x t The dilated convolution with a dilation rate of d is shown in Equation 12.
[0095]
[0096] Residual connections solve the problems of exploding and vanishing gradients caused by network deepening. Residual connections directly add the input to the output of each layer. The output formula is shown in Equation 13. o is the output of the residual module, X represents the input data, and F(X) represents the data after nonlinear transformation.
[0097] o=ReLU(X+F(X)) (13)
[0098] The BiTCN causal dilated convolution used in this paper has four layers, a kernel size of 2, and dilation coefficients d of 1, 2, 4, and 8 for each layer, respectively. This allows for a deeper model layer to achieve a higher and wider field of view. BiTCN's bidirectional structure allows the model to extract data features from both temporal dimensions, more accurately capturing local variations and dependencies in sequence information.
[0099] Step 4: The bidirectional temporal feature extraction module will continue to process the features obtained from the bidirectional local feature extraction module to capture the complex relationship between the previous and next moments of the feature.
[0100] The bidirectional temporal feature extraction module continues to process the features obtained from the bidirectional local feature extraction module, capturing long-term dependencies between features. This module is implemented using a BiGRU network, which consists of two unidirectional GRUs operating in opposite directions. The GRU is an improvement on the LSTM network, boasting fewer parameters and a simpler structure. While the LSTM has an input gate, an output gate, and a forget gate, the GRU network is trained using two gated units: an update gate and a reset gate. This allows the GRU to inherit the advantages of the LSTM while also enabling faster training. The GRU's update gate determines when to update the hidden state, as shown in Equation 5.
[0101] z t =σ(W z ·[h t-1 ,x t ]+b z ) (5)
[0102] The reset gate is responsible for determining when to allow past information to affect the current hidden state, as shown in Equation 6.
[0103] r t =σ(W r ·[h t-1 ,x t ]+b r ) (6)
[0104] The new candidate state formula is shown in Equation 7.
[0105]
[0106] The formula for updating the hidden state is as follows:
[0107]
[0108] Where x t Represents input data, h t is the output of the GRU unit, r t is the reset gate, z t is the update gate, z t With r t Joint control from h t-1 to h t The calculation of the hidden state, the update gate controls both the current input data and the previous memory information h t-1 , output a value z between 0 and 1 t , z t Decide to what extent h t-1 Pass to the next state. σ is the Sigmoid function, Wz , W r , W h are the update gate, reset gate and weight matrix of candidate hidden state respectively. Reset gate controls h t-1 For the result h t The importance of h t-1 When the correlation with the new memory is high, the reset gate can be used to enhance the influence of the previous memory. According to the calculation results of the reset gate, update gate and implicit state, the output h at the current moment can be obtained by the formula t .
[0109] The neuron state of the GRU neural network is transmitted layer by layer from front to back, without considering the dependency of data from back to front. The present invention selects a bidirectional gated recurrent unit BiGRU. BiGRU mines information on input features in both sequential and reverse order in the time dimension, extracts the dependency of features between the previous and next moments, and then adds the outputs of GRU units at different moments to eventually become the output layer of BiGRU. The GRU output node at each moment contains the complete forward and backward time series information of the current moment in the input sequence. Moreover, the existing methods for predicting SIE only extract the impact of forward data on backward data, and do not consider the complex dependency between backward data and forward data, resulting in poor modeling ability for time series dependency in the data. Therefore, the present invention uses BiGRU to simultaneously capture the complex relationship between the previous and next moments of the feature, and has a more accurate prediction of abnormal changes in sea ice during the melting season. BiGRU consists of two unidirectional GRUs in opposite directions. The forward GRU captures past information, and the backward GRU captures future information. The two GRUs jointly determine the output.
[0110] Step 5: The feature focus capture mechanism reduces the interference of irrelevant information by giving higher weight to relevant information.
[0111] Nine atmospheric and meteorological variables, as well as sea ice concentration, have varying degrees of influence on sea ice extent. To maximize the impact of the most strongly correlated variables on sea ice extent and minimize the impact of the least correlated variables, this paper introduces a feature-focused capture mechanism. This mechanism enables the model to automatically learn key information from the input sequence. By assigning higher weights to relevant information, the model can more effectively focus on components more relevant to sea ice extent and reduce interference from irrelevant information. Furthermore, while the aforementioned CNN can perform feature extraction and dimensionality reduction on time series, it cannot consider the magnitude of the influencing factors of each feature to select the most useful features. The BiGRU simultaneously mines input features in both sequential and reverse order along the temporal dimension, but it also suffers from a common drawback of RNNs: when processing long sequences, it cannot fully capture important information at each time step because the hidden state is transferred across time steps, and information gradually weakens or disappears over time. Therefore, the introduction of a feature-focused capture mechanism effectively addresses this issue by dynamically assigning weights to input features, thereby more accurately capturing important information at each time step.
[0112] The feature-focused capture mechanism is implemented through the attention mechanism. The basic idea of the attention mechanism is to filter out useful information through certain means, identifying factors that are more important to the prediction target, thereby improving prediction efficiency and reliability. The weight calculation formulas for the attention mechanism layer are shown in Equations 9, 10, and 11.
[0113] u t =tanh(W t h t +b t ) (9)
[0114]
[0115] l t =Σa t h t (12)
[0116]
[0117] Among them, W t represents the weight matrix, b t represents the bias coefficient, h t represents the output of the BiGRU layer, a t Represents the probability distribution value of the BiGRU layer output, that is, the attention weight, l t Represents the output of the attention layer.
[0118] In order to verify the effectiveness and advancement of the present invention, a comparative test was conducted between the present invention and five existing sea ice extent prediction algorithms. The experiments showed that the sea ice extent prediction accuracy of the algorithm of the present invention reached 99.58%, especially the prediction accuracy in the melting season reached 0.9934, and its prediction performance of sea ice extent in September and the freezing season was also excellent, which fully demonstrated the effectiveness and advancement of the algorithm.
[0119] It should be noted that, in the present invention patent, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.
[0120] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as within the scope of protection of the present invention.
Claims
1. A hybrid model sea ice extent prediction method based on a feature-focused capture mechanism, characterized by: The following steps are involved: S1. Obtain monthly EAR5 reanalysis data on sea ice concentration and atmospheric and meteorological variables as experimental data, and obtain monthly sea ice extent data from the National Snow and Ice Data Center as reference data. Preprocess the experimental and reference data to obtain the required standard data. S2. Use a spatial feature extraction module to extract the spatial features hidden in the two-dimensional sea ice concentration data, and fuse the extracted features with subsequent features as important predictors to predict future sea ice extent. The spatial feature extraction module is implemented using a convolutional neural network and is specifically designed as follows: The convolutional neural network of the spatial feature extraction module includes a four-convolutional layer block structure. Each convolutional layer block includes a convolution layer, a weight normalization layer, a ReLU layer, and a maximum pooling layer. The convolution kernel size of the convolution layer is 3×3. The convolution layer identifies spatial features in the sea ice concentration data through convolution operations. The maximum pooling layer performs a pooling operation with a 2×2 window with a stride of 2 to reduce the dimension of the features while retaining the main features. The data format of the input convolutional neural network is (10, 1, 1440, 261), where 10 is the batch size, 1 is the number of channels, and 1440 × 261 represents the latitude and longitude of the data. Multi-level feature extraction and dimensionality reduction are performed on the data, and the output feature shape is (10, 64, 90, 16). The obtained feature representation is passed to two fully connected layers for further processing to output spatial features, and the final output feature shape is (10,64); S3: Fuse the spatial features extracted in S2 with the atmospheric and meteorological multivariate features, and input the fused tensor into a bidirectional local feature extraction module to capture the bidirectional relationships and dependencies that only exist in a specific time or region. The bidirectional local feature extraction module is implemented using a bidirectional temporal convolutional network, and its specific design is as follows: The bidirectional temporal convolutional network consists of two components: forward temporal convolution and reverse temporal convolution. The forward temporal convolution and reverse temporal convolution extract features of different scales by stacking multiple convolutional layers, thereby capturing features of different time ranges in sequence data. By utilizing the multi-scale information extraction capability, TCN becomes more sensitive to local dependencies in sequence data and effectively extracts local features of time series. The TCN consists of three parts: causal convolution, dilated convolution, and residual connection. Causal convolution and dilated convolution are combined to form causal dilated convolution. The causal dilated convolution is defined as: filter F =( f 1, f 2,…, f K ), time series X =( x 1, x 2,…, x T ),exist x t The dilated convolution formula with a dilation rate of d is as follows: (3) The residual network is used to solve the problems of gradient explosion and gradient disappearance caused by network deepening. It directly adds the input of each layer to the output. The output formula is as follows: (4) in, o Represents the output of the residual module; X Represents input data; F ( X ) represents the data after nonlinear transformation; The causal dilation convolution is set to 4 layers in total, the kernel is set to 2, and the dilation coefficient d of each layer is set to 1, 2, 4, and 8 respectively; S4. Use the bidirectional temporal feature extraction module to continue processing the features extracted by the bidirectional local feature extraction module in S3 to capture the complex relationship between the time before and after the features. The bidirectional temporal feature extraction module is implemented by a bidirectional gated recurrent unit, and its specific design is as follows: The bidirectional gated recurrent unit is composed of two unidirectional GRUs in opposite directions; the GRU network is trained by two gating units: the update gate and the reset gate; The function of the update gate is expressed as follows: (5) The function of the reset gate is expressed as follows: (6) The new candidate state formula is as follows: (7) The formula for updating the hidden state is as follows: (8) in, x t Represents input data; h t Represents the output of the GRU unit; r t Reset gate. z t represents the update gate; σ Represents the Sigmoid function; W z , W r , W h Represent the weight matrices of the update gate, reset gate, and candidate hidden states respectively; S5. Design a feature-focused capture mechanism to give higher weight to relevant information and reduce interference from irrelevant information. The feature-focused capture mechanism is implemented through an attention mechanism. The feature-focused capture mechanism is used to enable the model to automatically learn key information in the input sequence. By giving higher weight to relevant information, the model pays more attention to the parts related to the sea ice range and reduces interference from irrelevant information.
2. The hybrid model sea ice extent prediction method based on the feature-focused capture mechanism according to claim 1 is characterized in that: For the experimental data and reference data obtained in S1, missing values are replaced using cubic spline interpolation, which is defined as follows: For n+1 known data points ( x 0, y 0)( x 1, y 1),…,( x n , y n ), a cubic polynomial between each adjacent data point S i ( x ) is expressed as: S i ( x )= a i + b i ( x - x i )+ c i ( x - x i ) 2 + d i ( x - x i ) 3 , and meet the following conditions: (1) S i ( x ) is a cubic polynomial in each small interval; (2) S i ( x ) has the same first-order and second-order derivatives at the connection between two adjacent small intervals, that is, ; (3) S i ( x ) passes through all given data points, i.e. S i ( x i )= y i ; (4) The boundary conditions are and ; Based on the above conditions, establish a linear equation system to solve the coefficients a i , b i , c i , d i ; After solving, when interpolating at a given point x, determine the interval where x is located [ x i , x i+1 ], using the corresponding cubic polynomial S i ( x ) to calculate and obtain the corresponding interpolation value.
3. The hybrid model sea ice extent prediction method based on the feature-focused capture mechanism according to claim 2 is characterized in that: The experimental data and reference data are preprocessed as described in S1 to obtain the required standard data, which specifically includes the following: The data is standardized using the MinMax standardization technique. The MinMax standardization formula is as follows: (1) in, Data Represents the original data; Data min Indicates the minimum value in the original data; Data max Indicates the maximum value in the original data; After the model predicts the sea ice extent data, the data is denormalized. The denormalization formula is as follows: Data * =( Data max - Data min ) + Data min (2) in, Data * represents the denormalized value, Represents the model prediction value.
