Sea ice multi-element combined forecasting method and system based on causal sharing

Abstract

The invention relates to a causal sharing-based sea ice multi-element combined forecasting method and system, and the method comprises the steps: extracting causal relationship values of different types of sea ice element data, and outputting a first feature map of the corresponding types of sea ice element data through respective branch networks, thereby constructing a causal relationship feature map; and then the causal relationship feature map is input to the corresponding branch network to realize information sharing and interaction between every two branch networks, and the causal relationship value is a two-dimensional space field between every two kinds of sea ice element data and contains physical constraints between the two kinds of sea ice element data, so that the difference between sea ice element independent forecasting models can be overcome, and the prediction accuracy is improved. And unification of multi-element forecasting results is realized.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of marine data processing, and in particular, to a joint forecasting method and system for multiple sea ice elements based on causal sharing. Background Art

[0002] Since this century, the Arctic sea ice cover has shown an obvious downward trend, and the development and utilization of the Arctic have gradually become a global hot issue. Accurately grasping the changing rules of sea ice is the key issue for the development and utilization of the Arctic, which puts forward higher requirements for the current forecasting ability of Arctic sea ice. However, current researchers still lack understanding of the physical laws of sea ice changes. The numerical forecasting of sea ice mainly considers the thermodynamic laws affecting sea ice changes. For some dynamic factors, due to the lack of clear physical laws, the parameterization schemes are deficient.

[0003] With the development of technologies such as satellite remote sensing and data assimilation, a large amount of remote sensing observation data and reanalysis data have been accumulated, providing a solid data foundation for the application of statistical methods in sea ice forecasting. And the wide application of deep learning methods has confirmed their excellent non-linear fitting ability, providing a reliable tool for Arctic sea ice forecasting. Currently, deep learning models have achieved good application effects in the medium- and long-term and short-term forecasting of sea ice elements. For example, the application of three main models, namely long short-term memory network (LSTM), convolutional neural networks (CNNs), and convolutional long short-term memory network (ConvLSTM), in the monthly-scale and daily-scale forecasting of sea ice concentration.

[0004] However, from the existing research, the current application of deep learning in sea ice forecasting mostly focuses on single-task forecasting of a certain physical feature of sea ice. Separately forecasting the physical features of sea ice easily causes inconsistencies in the physical laws among the physical features of sea ice. For example, grid points with zero sea ice concentration have non-zero sea ice thickness, which will restrict the accuracy and predictability of sea ice forecasting. Summary of the Invention

[0005] The following is an overview of the subject matter described in detail in this article. This overview is not intended to limit the scope of protection of the claims.

[0006] The main purpose of the embodiments of the present invention is to propose a joint forecasting method and system for multiple sea ice elements based on causal sharing, which can overcome the differences between single-forecasting models of sea ice elements and achieve the unification of multi-element forecasting results.

[0007] To achieve the above object, the first aspect of the embodiments of the present invention provides a joint forecasting method for multiple sea ice elements based on causal sharing, and the method includes: Obtain n different sea ice element data, and divide the n groups of different sea ice element data into s sea ice element data groups; each sea ice element data group includes 2 different sea ice element data, and any two of the n different sea ice element data form a sea ice element data group; Extract the two-dimensional spatial field between the time series corresponding to every two different sea ice element data as the causal relationship value according to the causal information flow theory; Construct a joint prediction model and train the joint prediction model; the joint prediction model includes s causal relationship modules and n branch networks with the same structure, wherein training the joint prediction model includes: Input the n different sea ice element data into the corresponding one of the branch networks respectively to obtain the first feature map corresponding to one kind of sea ice element data extracted by each branch network; Extract the causal relationship feature map corresponding to a group of sea ice element data groups through each causal relationship module according to the corresponding first feature map and the causal relationship value; Input the causal relationship feature maps into the two branch networks corresponding to a group of sea ice element data groups respectively, so that the branch networks generate the sea ice element prediction results corresponding to one kind of the sea ice element data according to the corresponding first feature map and the causal relationship feature map; Update the network parameters of the branch networks according to the sea ice element prediction results until the trained joint prediction model is obtained; Perform joint prediction on the n different target sea ice element data according to the trained joint prediction model.

[0008] A joint prediction method for multiple sea ice elements based on causal sharing provided by the present application has at least the following beneficial effects: By extracting the causal relationship value of different types of sea ice element data and constructing the causal relationship feature map with the first feature map output by the respective branch networks corresponding to the corresponding type of sea ice element data, and then inputting the causal relationship feature map into the corresponding branch network, information sharing and interaction between every two branch networks can be realized. Moreover, because the causal relationship value is the two-dimensional spatial field between every two sea ice element data and contains the physical constraints between each other, the differences between the single prediction models of sea ice elements can be overcome, and the unity of the multi-element prediction results can be realized.

[0009] In some embodiments, the causal relationship value between every two different sea ice element data is extracted by the following formula: ; wherein, represents the time series of a kind of sea ice element data For the time series of another sea ice element data of the causal relationship value, denotes determinant calculation, denotes covariance matrix, denotes the number of variables, is a variable, denotes the cofactor of, denotes all and the sample covariance between the Euler forward difference approximation of, denotes and the sample covariance between, denotes the sample variance of.

[0010] In some embodiments, the joint prediction model further includes a CBAM layer, and an output end of the CBAM layer is connected to an input end of each of the causal relationship modules; Before extracting the causal relationship feature map of a corresponding set of sea ice element data groups according to the corresponding first feature map and the causal relationship value, it further includes: Inputting the n first feature maps into the CBAM layer to enhance the attention mechanism of each of the n first feature maps through the CBAM layer, and obtaining n enhanced first feature maps; The extracting the causal relationship feature map of a corresponding set of sea ice element data groups according to the corresponding first feature map and the causal relationship value includes: Extracting the causal relationship feature map of a corresponding set of sea ice element data groups according to the corresponding enhanced first feature map and the causal relationship value.

[0011] In some embodiments, the causal relationship feature map of a set of sea ice element data groups includes: a first causal relationship feature map corresponding to the first sea ice element data in a set of sea ice element data groups and a second causal relationship feature map corresponding to the second sea ice element data; The extracting the causal relationship feature map of a corresponding set of sea ice element data groups according to the corresponding enhanced first feature map and the causal relationship value includes: Multiplying the first enhanced first feature map corresponding to the first sea ice element data by the first causal relationship value to obtain a first intermediate feature map, and adding the first intermediate feature map to the first enhanced first feature map to obtain the first causal relationship feature map; the first causal relationship value is the causal relationship value of the second sea ice element data to the first sea ice element data; Multiply the second enhanced first feature map corresponding to the second sea ice element data by the second causal relationship value to obtain a second intermediate feature map, and add the second intermediate feature map to the second enhanced first feature map to obtain the second causal relationship feature map; the second causal relationship value is the causal relationship value of the first sea ice element data to the second sea ice element data; Input the causal relationship feature map into two branch networks corresponding to a group of sea ice element data groups respectively, including: Input the first causal relationship feature map into one of the branch networks that extracts the first first feature map; Input the second causal relationship feature map into one of the branch networks that extracts the second first feature map.

[0012] In some embodiments, the network structure of any branch network includes an input layer, an encoder, a decoder, and an output layer connected in sequence; Both the input layer and the output layer are 1 convolutional layer; The encoder includes convolutional layers and downsampling layers, = 2 ; the first layer and the last layer of the encoder are both convolutional layers, and two consecutive convolutional layers are connected between every two downsampling layers; The decoder includes convolutional layers and upsampling layers; the first layer and the last layer of the decoder are both convolutional layers, and two consecutive convolutional layers are connected between every two upsampling layers; Inputting n different sea ice element data into one corresponding branch network respectively to obtain the first feature map corresponding to one kind of sea ice element data extracted by each branch network includes: Input each kind of sea ice element data into one corresponding branch network respectively to obtain the first feature map output by the th downsampling layer or the th upsampling layer of the corresponding branch network; is a positive integer less than or equal to , is a positive integer less than or equal to ; Inputting the causal relationship feature map into two branch networks corresponding to a group of sea ice element data groups respectively includes: Fuse the first feature maps output from the th downsampling layer in the two branch networks that respectively fuse the causal relationship feature map and a corresponding set of sea ice element data sets, to obtain two corresponding fused feature maps, and input each fused feature map into the first convolutional layer of a corresponding one of the branch networks; the first convolutional layer is a convolutional layer that is after and connected to the th downsampling layer; Or, Fuse the first feature maps output from the th upsampling layer in the two branch networks that respectively fuse the causal relationship feature map and a corresponding set of sea ice element data sets, to obtain two corresponding fused feature maps, and input each fused feature map into the first convolutional layer of a corresponding one of the branch networks; the first convolutional layer is a convolutional layer that is after and connected to the th upsampling layer.

[0013] In some embodiments, the output end of at least one convolutional layer in the encoder is connected to the input end of a convolutional layer with the same feature map dimension in the decoder by a residual connection.

[0014] In some embodiments, n is 3, and the n different sea ice element data include sea ice concentration data, sea ice velocity data, and sea ice thickness data.

[0015] To achieve the above object, a second aspect of the embodiments of the present invention provides a joint forecasting system for multi-elements of sea ice based on causal sharing, and the system includes: A data acquisition unit, configured to acquire n different sea ice element data, and divide the n groups of different sea ice element data into s sea ice element data sets; each sea ice element data set includes 2 different sea ice element data, and any two of the n different sea ice element data form a sea ice element data set; A relationship value acquisition unit, configured to extract a two-dimensional spatial field between the time series corresponding to every two different sea ice element data as a causal relationship value according to the causal information flow theory; A model training unit, configured to construct a joint forecasting model and train the joint forecasting model; the joint forecasting model includes s causal relationship modules and n branch networks with the same structure, and training the joint forecasting model includes: Input the n different sea ice element data into a corresponding one of the branch networks respectively, to obtain the first feature maps of a corresponding one of the sea ice element data extracted by each branch network; Each of the causal relationship modules extracts a causal relationship feature map of a corresponding set of sea ice element data groups according to the corresponding first feature map and the causal relationship value; Input the causal relationship feature maps into the two branch networks of the corresponding set of sea ice element data groups respectively, so that the branch networks generate sea ice element forecast results corresponding to one type of the sea ice element data according to the corresponding first feature map and the causal relationship feature map; Gradient update the network parameters of the branch networks according to the sea ice element forecast results until a trained joint forecast model is obtained; A model application unit for jointly forecasting n different target sea ice element data according to the trained joint forecast model.

[0016] To achieve the above object, a third aspect of the embodiments of the present invention provides an electronic device, including: at least one control processor and a memory for communicatively connecting with the at least one control processor; the memory stores instructions executable by the at least one control processor, and the instructions are executed by the at least one control processor so that the at least one control processor can execute the above-mentioned joint forecasting method for multiple sea ice elements based on causal sharing.

[0017] To achieve the above object, a fourth aspect of the embodiments of the present invention provides a computer-readable storage medium, the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to make a computer execute the above-mentioned joint forecasting method for multiple sea ice elements based on causal sharing.

[0018] It can be understood that the beneficial effects of the above-mentioned second aspect to the fourth aspect compared with the related technologies are the same as those of the first aspect compared with the related technologies. For the relevant descriptions, reference can be made to the relevant descriptions in the first aspect, and details will not be repeated here. Description of the Drawings

[0019] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for use in the embodiments or the related technologies. Obviously, the drawings in the following description are only some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0020] Figure 1 It is a schematic diagram of an embodiment of a joint forecasting method for multiple sea ice elements based on causal sharing provided by the present application; Figure 2 It is a schematic diagram of an embodiment of the causal sharing module Cause-AM provided by the present application; Figure 3 It is a schematic diagram of an embodiment of the branch network provided by this application; Figure 4 It is a schematic diagram of the data interaction between the branch network and Cause-AM provided by an embodiment of this application; Figure 5 It is a schematic diagram of an embodiment of the joint prediction model provided by this application; Figure 6 It is a schematic diagram of an embodiment of the joint prediction system for multi-element sea ice based on causal sharing provided by this application; Figure 7 It is a schematic diagram of an embodiment of the electronic device provided by this application. Detailed implementation manners

[0021] In order to make the objectives, technical solutions and advantages of this application clearer, the following further details this application in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application.

[0022] As Figure 1 , an embodiment of this application provides a joint prediction method for multi-element sea ice based on causal sharing. The method includes steps S110 to S140: Step S110, obtain n types of different sea ice element data, and divide the n groups of different sea ice element data into s sea ice element data groups; each sea ice element data group includes 2 types of different sea ice element data, and any two types of sea ice element data among the n types of different sea ice element data form a sea ice element data group.

[0023] Step S120, extract the two-dimensional spatial field between the time series corresponding to every two types of different sea ice element data as the causal relationship value according to the causal information flow theory.

[0024] Step S130, construct a joint prediction model and train the joint prediction model; the joint prediction model includes s causal relationship modules and n branch networks with the same structure. Among them, training the joint prediction model includes: Input the n types of different sea ice element data into the corresponding one branch network respectively to obtain the first feature map corresponding to one type of sea ice element data extracted by each branch network; Through each causal relationship module, extract the causal relationship feature map corresponding to a group of sea ice element data according to the corresponding first feature map and the causal relationship value; Input the causal relationship feature maps into the two branch networks corresponding to a group of sea ice element data respectively, so that the branch network generates the sea ice element prediction result corresponding to one type of sea ice element data according to the corresponding first feature map and the causal relationship feature map; Gradient update is performed on the network parameters of the branch network according to the prediction results of sea ice elements until a trained joint prediction model is obtained.

[0025] Step S140: Perform joint prediction on n different target sea ice element data according to the trained joint prediction model.

[0026] In this embodiment, steps S110 to S120 are the processes of processing the training set for training the joint prediction model; step S130 is the process of model training, and step S140 is the application process.

[0027] In step S110, n is greater than or equal to 3. Taking 3 as an example, the sea ice element data can be sea ice concentration data, sea ice velocity data, and sea ice thickness data. The quantitative relationship between n and s is: , is the factorial. If n = 3, then s = 3; if n = 4, then s = 6.

[0028] Here, s sea ice element data groups are divided to facilitate the subsequent model to perform interactive processing on the characteristics of the data groups.

[0029] For example, if the sea ice element data is sea ice concentration data, sea ice velocity data, and sea ice thickness data, the data sources can be: the fifth-generation reanalysis product of the European Centre for Medium-Range Weather Forecasts, the product of the National Snow and Ice Data Center of the United States, and the Pan-Arctic Ocean Simulation and Assimilation System. Specific descriptions will be given in subsequent embodiments.

[0030] In step S120, taking sea ice concentration data, sea ice velocity data, and sea ice thickness data as examples, this step calculates the causal relationship among the three factors of sea ice concentration data, sea ice velocity data, and sea ice thickness data. Let the time series of two factors be and , then For , the calculation formula for the causal effect on ; Among them, represents 's causal effect on , represents determinant calculation, represents the covariance matrix, represents the number of variables, represents 's cofactor, represents all and ( is time) the sample covariance between the Euler forward difference approximation of denote and the sample covariance between denote the sample variance of

[0031] On the contrary for the calculation formula of the causal effect is as follows ; Since the sea ice element data all have spatial grids, the causal relationship values are calculated based on the time series of each grid point of the spatial grid

[0032] In step S130, first, the joint prediction model includes multiple branch networks. The branch networks are arranged in parallel and are independent of each other. It should be noted that the independence here only refers to the independent network structures, and the data (in the form of features) will flow among multiple branch networks. The number of branch networks is the same as the number of types of sea ice element data, that is, one branch network processes one type of sea ice element data alone. The model also includes multiple causal relationship modules. The number of causal relationship modules is the same as the number of groups of sea ice element data. The main function of the causal relationship module is to construct the data interaction between every two branch networks. See the subsequent description for details

[0033] The training of the model includes the following steps 1) Input n different types of sea ice element data into the corresponding one branch network respectively to obtain the first feature map of one type of sea ice element data extracted by each branch network

[0034] The branch network can be Unet or ConvLSTM. This embodiment does not make specific limitations

[0035] Because each branch network focuses on the processing of one type of sea ice element data, the first feature map extracted by each branch network is also the feature map extracted from one type of sea ice element data

[0036] 2) Through each causal relationship module, extract the causal relationship feature map of the corresponding group of sea ice element data according to the corresponding first feature map and causal relationship value

[0037] Taking the sea ice concentration data and sea ice velocity data as an example, the corresponding causal relationship module of this group is based on the first feature output by branch network 1 (the branch network processing sea ice concentration data) Figure 1 and the first feature output by branch network 2 (the branch network processing sea ice velocity data) Figure 2, as well as the causal relationship value 1 of sea ice concentration data to sea ice velocity data and the causal relationship value 2 of sea ice velocity data to sea ice concentration data, extract the causal relationship feature map of this group of sea ice concentration data and sea ice velocity data.

[0038] 3) Input the causal relationship feature maps into the two branch networks corresponding to a group of sea ice element data respectively, so that the branch networks generate the sea ice element prediction results corresponding to a type of sea ice element data according to the corresponding first feature map and the causal relationship feature map.

[0039] Following the above embodiment, assume that the causal relationship feature map of this group of sea ice concentration data and sea ice velocity data is extracted. Here, the causal relationship feature maps will be input into branch network 1 and branch network 2 of this group of sea ice concentration data and sea ice velocity data respectively to construct the information interaction between branch network 1 and branch network 2.

[0040] In this embodiment, by extracting the causal relationship values of different types of sea ice element data and constructing the causal relationship feature map with the first feature maps output by their respective branch networks corresponding to the types of sea ice element data, and then inputting the causal relationship feature map into the corresponding branch network, the information sharing and interaction between every two branch networks are realized. Moreover, because the causal relationship value is a two-dimensional spatial field between every two types of sea ice element data, which contains the physical constraints between them, this can overcome the differences between the single prediction models of sea ice elements and achieve the unification of multi-element prediction results. Avoid problems such as the reduction of the usability of prediction data caused by large differences between different elements.

[0041] In some embodiments of the present application, the joint prediction model further includes a CBAM layer, and the output end of the CBAM layer is connected to the input end of each causal relationship module; Before extracting the causal relationship feature map of a group of sea ice element data according to the corresponding first feature map and causal relationship value in step S130 above, the method further includes: Input the n first feature maps into the CBAM layer to enhance the n first feature maps respectively through the attention mechanism of the CBAM layer, and obtain n enhanced first feature maps; The extraction of the causal relationship feature map of a group of sea ice element data according to the corresponding first feature map and causal relationship value in step S130 above includes: Extract the causal relationship feature map of a group of sea ice element data according to the corresponding enhanced first feature map and causal relationship value.

[0042] In this embodiment, CBAM consists of a Channel Attention Module and a Spatial Attention Module.

[0043] CBAM can enhance each of the n first feature maps one by one, which can improve the performance of the model in capturing features and its generalization ability.

[0044] After CBAM enhancement, n enhanced first feature maps are obtained. Then, according to the corresponding enhanced first feature maps and causal relationship values, causal relationship feature maps of a corresponding set of sea ice element data groups are extracted.

[0045] In some embodiments of the present application, the causal relationship feature maps of a set of sea ice element data groups include: a first causal relationship feature map corresponding to the first sea ice element data in a set of sea ice element data groups and a second causal relationship feature map corresponding to the second sea ice element data.

[0046] The steps of S130 above: extracting the causal relationship feature maps of a corresponding set of sea ice element data groups according to the corresponding enhanced first feature maps and causal relationship values, include: Multiply the first enhanced first feature map corresponding to the first sea ice element data by the first causal relationship value to obtain a first intermediate feature map, and add the first intermediate feature map to the first enhanced first feature map to obtain the first causal relationship feature map; the first causal relationship value is the causal relationship value of the second sea ice element data to the first sea ice element data; Multiply the second enhanced first feature map corresponding to the second sea ice element data by the second causal relationship value to obtain a second intermediate feature map, and add the second intermediate feature map to the second enhanced first feature map to obtain the second causal relationship feature map; the second causal relationship value is the causal relationship value of the first sea ice element data to the second sea ice element data.

[0047] The steps of S130 above: inputting the causal relationship feature maps into two branch networks of a corresponding set of sea ice element data groups respectively, include: Input the first causal relationship feature map into a branch network that extracts the first first feature map; Input the second causal relationship feature map into a branch network that extracts the second first feature map.

[0048] In this embodiment, taking sea ice concentration data and sea ice velocity data as examples, assume that The output of the branch network 1 corresponding to the sea ice concentration data is the first feature map f1, and the first enhanced first feature map is F1. The output of the branch network 2 corresponding to the sea ice velocity data is the first feature map f2, and the second enhanced first feature map is F2. The first causality value (the causality value of the sea ice velocity data on the sea ice concentration data) is S1, and the second causality value (the causality value of the sea ice concentration data on the sea ice velocity data) is S2.

[0049] Take S1 as the weight of F1, that is, S1 * F1 is the first intermediate feature map; take S2 as the weight, that is, S2 * F2 is the second intermediate feature map; then, add the first intermediate feature map and the first enhanced first feature map, that is, S1 * F1 + F1 is the first causality feature map, add the second intermediate feature map and the second enhanced first feature map, that is, S2 * F2 + F2 is the second causality feature map, and finally input S1 * F1 + F1 into the branch network 1, and input S2 * F2 + F2 into the branch network 2.

[0050] In this embodiment, the causality value is used as the weight of the enhanced first feature map for multiplication to construct the interaction between the information of the two types of sea ice data. In addition, the multiplication result is added to the enhanced first feature map to improve the information volume in the feature map.

[0051] In some embodiments of the present application, the network structure of any branch network includes an input layer, an encoder, a decoder, and an output layer connected in sequence; Both the input layer and the output layer are 1 convolutional layer; The encoder includes convolutional layers and downsampling layers, = 2 ; the first layer and the last layer of the encoder are both convolutional layers, and two connected convolutional layers are connected between every two downsampling layers; The decoder includes convolutional layers and upsampling layers; the first layer and the last layer of the decoder are both convolutional layers, and two connected convolutional layers are connected between every two upsampling layers; Input n different types of sea ice element data into the corresponding one branch network respectively to obtain the first feature map corresponding to one type of sea ice element data extracted by each branch network, including: Input each type of sea ice element data into the corresponding one branch network respectively to obtain the first feature map output by the th downsampling layer or the th upsampling layer of the corresponding one branch network; is less than or equal to a positive integer, is less than or equal to a positive integer; Input the causal relationship feature map into two branch networks corresponding to a set of sea ice element data groups respectively, including: Fuse the causal relationship feature map and the first feature map output by the th downsampling layer in two branch networks corresponding to a set of sea ice element data groups respectively to obtain two corresponding fused feature maps, and input each fused feature map into the first convolutional layer of the corresponding branch network; The first convolutional layer is the convolutional layer after and connected to the th downsampling layer; Or, Fuse the causal relationship feature map and the first feature map output by the th upsampling layer in two branch networks corresponding to a set of sea ice element data groups respectively to obtain two corresponding fused feature maps, and input each fused feature map into the first convolutional layer of the corresponding branch network; The first convolutional layer is the convolutional layer after and connected to the th upsampling layer.

[0052] In this embodiment, the specific structure of the branch network is set. In the branch network, each downsampling layer can generate a first feature map, and each upsampling layer can generate a first feature map. After generating the first feature map and obtaining the causal relationship feature map through the above processing, the first feature map generated by the downsampling layer can be combined and input into the convolutional layer after the downsampling layer, or the first feature map generated by the upsampling layer can be combined and input into the convolutional layer after the upsampling layer. The present application can perform multiple information interactions to enhance the ability of the model to extract information in features.

[0053] In some embodiments of the present application, the output end of at least one convolutional layer in the encoder is residually connected to the input end of the convolutional layer with the same feature map dimension in the decoder.

[0054] In this embodiment, the skip residual connection is added, which can prevent the gradient from vanishing and improve the information volume of the original skip connection, thereby improving the prediction accuracy.

[0055] For ease of understanding, as Figures 2 to 5 , the present application provides a joint prediction method for multiple sea ice elements based on causal sharing, specifically a joint intelligent prediction method for three elements of sea ice concentration, sea ice velocity, and sea ice thickness. The method includes the following steps: Step S910, obtain satellite remote sensing data or model data of sea ice concentration, sea ice velocity, and sea ice thickness to construct a training set and a validation set.

[0056] Obtain daily sea ice concentration data products, daily sea ice velocity data, and daily sea ice thickness data products. The daily sea ice concentration data products come from the fifth-generation reanalysis product (ERA5) of the European Centre for Medium-Range Weather Forecasts. The daily sea ice velocity data products come from the products of the National Snow and Ice Data Center (NSIDC) in the United States. The daily sea ice thickness data products come from the Pan-Arctic Ice Ocean Modeling and Assimilation System (PIOMAS).

[0057] The time span of the data is daily resolution data from 1980 to 2023, a total of 16,071 days. For the data of PIOMAS, the missing February data in leap years are filled with the previous day's data. The data cover the entire Arctic. To unify the data format, the data of NSIDC and PIOMAS are uniformly interpolated to the grid points of ERA5, and the spatial resolution is , the latitude ranges from to , so the spatial range of the final data is . The grid fields of each day are combined to construct a three-dimensional matrix with the scale of [longitude, latitude, time], and the symbol size is , that is , .

[0058] Specifically, preprocess the sea ice concentration data from ERA5. The data storage format is a NETCDF file. The preprocessing steps include: (1) Read the data information of sea ice concentration, including variable names, longitude, latitude, and the original sea ice concentration data in the file; (2) Select the longitude and latitude ranges of the entire Arctic, and discard the data outside the research scope. Therefore, the spatial range is ; (3) The missing value of the original sea ice concentration data is -32767. Assign the missing value of the original data as "Nan" as a marker.

[0059] Secondly, preprocess the sea ice velocity data from NSIDC. The data storage format is a NETCDF file. The preprocessing steps include: (1) Read the data information of sea ice velocity, including variable names, longitude, latitude, sea ice velocity u-direction data, and sea ice velocity v-direction data in the file; (2) Select the longitude and latitude ranges of the entire Arctic. Due to the projection method, the spatial range of the data is ; (3) The missing value of the sea ice velocity data is -9999. Assign the missing value of the data as "Nan" as a marker; (4)The projections in the u and v directions are transformed into the longitude and latitude directions. The projection formula for the sea ice velocity at the North Pole is as follows: ; ; Among them, represents the zonal sea ice velocity with the east direction being positive, represents the meridional sea ice velocity with the north direction being positive, represents the longitude position of the original sea ice velocity in the u and v directions.

[0060] Secondly, the spatial grid of the sea ice velocity is unified into the spatial grid of the sea ice concentration through bicubic interpolation. Let the coordinates of the interpolation point to be solved be , and given the data of 16 pixel coordinate points (grids) around it, it is also necessary to calculate the weights of each of the 16 points. Taking the pixel coordinate point as an example, the data of the interpolation point to be solved will be obtained through the following calculations: ; ; ; ; ; ; Among them, represents the data value of the corresponding grid point, and represent that the point is at a distance of and from the interpolation point to be solved in the and directions respectively, is the interpolation weight kernel. Similarly, the weights of the remaining 15 pixel coordinate points can be obtained.

[0061] Secondly, preprocess the sea ice thickness data from PIOMAS. The data storage format is the hiday.H file, and the longitude and latitude information storage format is the.dat file. The preprocessing steps include: (1)Read the data information of the sea ice thickness by byte count, reading with a float32 byte length; (2)Reshape the read byte data into a data grid; (3)Similarly, read the longitude and latitude information by byte and reshape it into ; Secondly, the spatial grid of sea ice thickness is unified into the spatial grid of sea ice concentration through bicubic interpolation. The interpolation formula is the same as that in the sea ice velocity part and will not be elaborated here. Secondly, the magnitudes of sea ice concentration, sea ice velocity, and sea ice thickness are unified. Among them, the magnitude of sea ice concentration is multiplied by 100 so that its scaled magnitude is between 0 and 100. The magnitude of sea ice velocity is multiplied by 2 so that its scaled magnitude is between -120 and 120. The magnitude of sea ice thickness is between 0 and 90.

[0062] Secondly, the "Nan" values in the sea ice concentration, sea ice velocity, and sea ice thickness data are processed. The parts marked with "Nan" values are uniformly assigned 0, and it is considered that there is no sea ice here from a physical meaning perspective.

[0063] Secondly, the training set and validation set are divided according to the preset forecast duration and the length of the time series. For example, the preset model forecasts the forecast data for the next 10 days based on the historical data of 10 days, and uses the data from 1980 to 2022 as the training set and the data in 2023 as the validation set. Among them, the single-day data of sea ice concentration occupies 1 channel. The sea ice velocity is represented separately by the meridional and zonal directions, so the single-day data occupies 2 channels. The single-day data of sea ice thickness occupies 1 channel. In the case of forecasting 10 days based on 10 days, the sea ice concentration of the input data occupies 10 channels, the sea ice velocity data occupies 20 channels, and the sea ice thickness data occupies 10 channels. The sea ice concentration of the output data occupies 10 channels, the sea ice velocity data occupies 20 channels, and the sea ice thickness data occupies 10 channels.

[0064] Step S920, calculate the causal relationship values between factors. Calculate the causality between sea ice concentration and sea ice velocity, sea ice concentration and sea ice thickness, and sea ice velocity and sea ice thickness respectively. Let the time series of two factors be

[0065] and and , then For the calculation formula of the causal effect on is as follows: Among them, represents the causal relationship value of on represents the determinant calculation, represents the covariance matrix, represents the number of variables, represents the cofactor of represents all and ( The sample covariance between the Euler forward difference approximation at time denotes and the sample covariance between denotes the sample variance of

[0066] Finally, three causal relationship values are obtained, all of which are two-dimensional spatial fields corresponding to the spatial size of sea ice elements .

[0067] Step S930, construct a joint prediction model

[0068] The joint prediction model in this embodiment is composed of a deep learning network (Cause-Unet) with three branches. Specifically The expression for processing by the joint prediction model is as follows ; where is the mapping function of Cause-Unet. The input of branch network 1 for sea ice concentration is and the output is The input of branch network 2 for sea ice velocity is and the output is The input of network 3 for sea ice thickness is and the output is . , and respectively represent the weights and biases of the three branch networks, represents the weights and biases of the causal sharing module Cause-AM. Among them, Cause-AM includes a CBAM layer and three causal relationship sub-modules

[0069] In this embodiment, three independent basic branches are constructed based on Unet, namely branch network 1, branch network, and branch network 3, which are respectively used for predicting sea ice concentration, sea ice velocity, and sea ice thickness. Each of the three branch networks includes an input layer, an encoder, a decoder, and an output layer. The input layer is a convolutional layer, the encoder includes 6 convolutional layers and 3 downsampling layers, the decoder includes 6 convolutional layers and 3 upsampling layers, and the output layer includes 1 convolutional layer. Residual connections connect the feature maps output after the first convolutional layer of the encoder and the first upsampling of the decoder, the feature maps output after the third convolutional layer of the encoder and the third upsampling of the decoder, and the feature maps output after the fifth convolutional layer of the encoder and the fifth upsampling of the decoder. The feature map output by each upsampling layer or downsampling layer is the first feature map

[0070] Taking branch network 1 as an example, the convolutional kernel size is all , the PReLU activation function is adopted, and the "same" padding method is used. The input layer inputs sea ice concentration data, with 64 hidden layers and a convolution kernel size of ; then comes the first layer of the encoder, with 64 hidden layers; then comes the downsampling layer, where the length and width of the spatial field are reduced by 1 / 2; then come the second and third layers of the encoder, with 128 hidden layers; then comes the downsampling layer, where the length and width of the spatial field are reduced by 1 / 2; then come the fourth and fifth layers of the encoder, with 256 hidden layers; then comes the downsampling layer, where the length and width of the spatial field are reduced by 1 / 2; then comes the sixth layer of the encoder, with 512 hidden layers; then comes the first layer of the decoder, with 512 hidden layers; then comes the upsampling layer, where the length and width of the spatial field are expanded by 1 times; then come the second and third layers of the decoder, with 256 hidden layers; then comes the upsampling layer, where the length and width of the spatial field are expanded by 1 times; then come the fourth and fifth layers of the decoder, with 128 hidden layers; then comes the upsampling layer, where the length and width of the spatial field are expanded by 1 times; then comes the sixth layer of the decoder, with 64 hidden layers; then comes the output layer, with 10 hidden layers.

[0071] Taking Branch Network 1 as an example, a residual connection is made to the feature map of Branch Network 1. The output feature map of the first layer of the encoder is processed by a convolutional layer with 128 convolutional kernels. The convolutional kernel size of the convolutional layer is 3×3. The PReLU activation function is adopted, and the "same" padding method is used. The output of the residual connection is added to the feature map output by the upsampling of the first layer of the decoder to obtain a new feature map.

[0072] Taking Branch Network 1 as an example, a bidirectional connection is established between the feature map of Branch Network 1 and the causal sharing layer. A bidirectional connection is established between the output of the downsampling layer of the encoder and the Cause-AM of the causal sharing layer. A bidirectional connection is established between the output of the upsampling layer of the encoder and the Cause-AM. Among them, a connection relationship is established between the feature map of Branch Network 1 and the causal relationship sub-module 1 and causal relationship sub-module 2 of the Cause-AM.

[0073] Based on Branch Network 1, Branch Network 2, Branch Network 3 and the causal relationship sub-module, a three-branch deep learning network Cause-Unet is constructed. Among them, the process of Branch Network 1 obtaining includes: ; ; ; ; ; ; ; Among them, represents the input sea ice concentration, represents multiple consecutive adjacent feature processing layers, , represents the downsampling layer, represents the upsampling layer, represents the causal relationship sub-module 1, represents the causal relationship sub-module 2, is the predicted sea ice concentration.

[0074] The process of obtaining by branch network 2 includes: ; ; ; ; ; ; ; Among them, represents the input sea ice velocity, represents multiple consecutive adjacent feature processing layers, , represents the downsampling layer, represents the upsampling layer, represents the causal relationship sub-module 1, represents the causal relationship sub-module 3, is the predicted sea ice velocity.

[0075] The process of obtaining by branch network 3 includes: ; ; ; ; ; ; ; Among them, represents the input original sea ice thickness, represents multiple consecutive adjacent feature processing layers, , represents the downsampling layer, Indicates the upsampling layer, Indicates the causal relationship sub-module 2, Indicates the causal relationship sub-module 3, Is the predicted sea ice thickness.

[0076] The three types of sub-modules represent the causal relationships of pairwise interactions between factors and can be calculated using the following method. Let the time series of two factors be and , then For The calculation formula for the causal effect is as follows: ; Among them, Indicates For The causal effect, Indicates the determinant calculation, Indicates the covariance matrix, Indicates the number of variables, Indicates The cofactor of, Indicates all And ( Is time) The sample covariance between the Euler forward difference approximations of, Indicates And The sample covariance between, Indicates The sample variance of.

[0077] The joint prediction model builds three branch networks and Cause-AM for sharing information between branches, realizing the unified prediction of sea ice concentration, sea ice velocity and sea ice thickness. In the case of predicting the next 10 days based on the historical 10 days, the sea ice concentration occupies 10 channels, the sea ice velocity is divided into longitude and latitude directions and occupies 20 channels, and the sea ice thickness occupies 10 channels.

[0078] It should be noted that the causal relationship sub-module in Cause-AM provides a function of selecting the corresponding causal relationship value and multiplying the causal relationship value as a weight by the enhanced first feature map. Since a causal relationship sub-module needs to process the first feature maps of the inputs of two branches, therefore, in the face of different first feature maps, the corresponding causal relationship value needs to be selected.

[0079] Step S940: Based on the data set obtained in S910, guided by the loss function proposed in S920, train and validate the joint prediction model obtained in S930.

[0080] For the loss function, loss functions such as mean absolute error and spatial similarity commonly used in deep learning can be adopted, or a loss function can be customized according to the task requirements.

[0081] Under the deep learning framework of pytorch, set the learning rate of the joint prediction model to 0.0001, the batch size for each batch to 16, train for a total of 20 epochs, decay the learning rate by 0.85 every 5 epochs, use the Adam optimizer for parameter tuning, train the model based on the training set obtained in step S910, and calculate the loss of each batch as the basis for parameter tuning in backpropagation.

[0082] For example, in this embodiment, the mean absolute error (MAE) is used to evaluate the prediction effect of each model. The calculation formula of MAE is as follows: ; Where represents the result predicted by the joint prediction model, represents the label data, represents the number of grid points. For example, the sea ice concentration, sea ice velocity, and sea ice thickness together have grid points.

[0083] Step S950: Use the verified joint prediction model to perform joint prediction on the target sea ice concentration data, target sea ice velocity data, and target sea ice thickness data, and obtain the daily prediction results with unified sea ice elements.

[0084] For the target sea ice concentration data, target sea ice velocity data, and target sea ice thickness data of the new non-training set, it is necessary to unify the grid points and magnitudes and process the missing data according to the preprocessing method in step S910.

[0085] Input the preprocessed target sea ice concentration data, target sea ice velocity data, and target sea ice thickness data into the joint prediction model in batches, and thus obtain the daily prediction results with unified sea ice concentration, sea ice velocity, and sea ice thickness.

[0086] The advantages of this embodiment are as follows: (1) This method introduces the causal information theory to calculate the pairwise causal relationships between the three elements of sea ice concentration, sea ice velocity, and sea ice thickness, providing constraints for information sharing and interaction between the three elements, and improving the interpretability of the model and the information exchange efficiency between branches.

[0087] (2) The three-branch network structure proposed in this method can simultaneously predict sea ice concentration, sea ice velocity, and sea ice thickness. Compared with training three independent deep learning models separately, it can reduce the overhead of model training and use, and has a higher integration degree, which is helpful for popularization and application.

[0088] (3) This method proposes Cause-AM to share and interact information among the three branch networks. While realizing the joint prediction of sea ice concentration, sea ice velocity, and sea ice thickness, it can ensure the unity of the prediction results of sea ice concentration, sea ice velocity, and sea ice thickness, and avoid problems such as the reduction of the availability of prediction data caused by large differences between different elements.

[0089] (4) The three-branch network structure proposed by this method has independent input ends. Therefore, it can simultaneously use sea ice concentration, sea ice velocity, and sea ice thickness data from different sources or structures. By setting the input layer parameters of each branch, the fusion of multi-source heterogeneous information can be quickly achieved, improving the applicability of the model.

[0090] An embodiment of this application also provides a joint prediction system for multi-elements of sea ice based on causal sharing. The system includes: a data acquisition unit, a relationship value acquisition unit, a model training unit, and a model application unit.

[0091] The data acquisition unit is used to acquire n types of different sea ice element data and divide the n groups of different sea ice element data into s sea ice element data groups; each sea ice element data group includes 2 types of different sea ice element data, and any two types of sea ice element data among the n types of different sea ice element data form a sea ice element data group. Both n and s are positive integers greater than 2. The relationship value acquisition unit is used to extract the two-dimensional spatial field between the time series corresponding to every two different sea ice element data as the causal relationship value according to the causal information flow theory. The model training unit is used to construct a joint prediction model and train the joint prediction model; the joint prediction model includes s causal relationship modules and n branch networks with the same structure. Among them, training the joint prediction model includes: Inputting the n types of different sea ice element data into the corresponding one branch network respectively to obtain the first feature map of the corresponding one type of sea ice element data extracted by each branch network; Extracting the causal relationship feature map of the corresponding group of sea ice element data through each causal relationship module according to the corresponding first feature map and causal relationship value; Inputting the causal relationship feature maps into the two branch networks of the corresponding group of sea ice element data respectively, so that the branch network generates the sea ice element prediction result of the corresponding one type of sea ice element data according to the corresponding first feature map and causal relationship feature map; Updating the network parameters of the branch network according to the sea ice element prediction result until the trained joint prediction model is obtained; The model application unit is used to perform joint prediction on the n types of different target sea ice element data according to the trained joint prediction model.

[0092] It should be noted that the joint forecasting system for multiple sea ice elements based on causal sharing provided in this embodiment and the above-mentioned joint forecasting method for multiple sea ice elements based on causal sharing are based on the same inventive concept. Therefore, the relevant content of the above-mentioned joint forecasting method for multiple sea ice elements based on causal sharing also applies to the content of the joint forecasting system for multiple sea ice elements based on causal sharing. Therefore, it will not be elaborated here.

[0093] Such as Figure 7 , this embodiment of the present application also provides an electronic device, which includes: At least one memory; At least one processor; At least one program; The program is stored in the memory, and the processor executes at least one program to implement the method described above in this disclosure.

[0094] The electronic device can be any intelligent terminal including a mobile phone, a tablet computer, a personal digital assistant (PDA), an in-vehicle computer, etc.

[0095] The electronic device of this embodiment of the present application will be introduced in detail below.

[0096] The processor 1600 can be implemented by using a general-purpose central processing unit (CPU), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, etc., and is used to execute relevant programs to implement the technical solutions provided by the embodiments of the present invention; The memory 1700 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM), etc. The memory 1700 can store an operating system and other application programs. When implementing the technical solutions provided in the embodiments of this specification through software or firmware, the relevant program codes are stored in the memory 1700 and are called by the processor 1600 to execute the joint forecasting method for multiple sea ice elements based on causal sharing in the embodiments of the present invention.

[0097] The input / output interface 1800 is used to implement information input and output; A communication interface 1900 is used to implement communication and interaction between this device and other devices. It can achieve communication through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.); A bus 2000 transmits information between various components of the device (such as a processor 1600, a memory 1700, an input / output interface 1800, and a communication interface 1900); Among them, the processor 1600, the memory 1700, the input / output interface 1800, and the communication interface 1900 are communicatively connected to each other inside the device through the bus 2000.

[0098] An embodiment of the present invention also provides a storage medium, which is a computer-readable storage medium. The computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions are used to cause a computer to execute the above method.

[0099] As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs and non-transitory computer-executable programs. In addition, the memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage devices. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories can be connected to the processor through a network. Examples of the above networks include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0100] The embodiments described in the present invention are for more clearly illustrating the technical solutions of the embodiments of the present invention, and do not constitute a limitation on the technical solutions provided by the embodiments of the present invention. Those skilled in the art can know that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of the present invention are equally applicable to similar technical problems.

[0101] Those skilled in the art can understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present invention, and may include more or fewer steps than those shown, or combine certain steps, or different steps.

[0102] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, that is, they may be located in one place, or may be distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0103] Those of ordinary skill in the art will understand that all or some of the steps in the methods disclosed above, and the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or a suitable combination thereof.

[0104] As used in the specification of this application and the above drawings, the terms "first", "second", "third", "fourth", etc. (if any) are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments of this application described here can be implemented in an order different from those illustrated or described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0105] It should be understood that in this application, "at least one (item)" means one or more, and "a plurality" means two or more. "And / or" is used to describe the relationship between related objects and indicates that there can be three relationships. For example, "A and / or B" can mean: only A exists, only B exists, and both A and B exist at the same time. Here, A and B can be singular or plural. The character " / " generally indicates that the related objects before and after are in an "or" relationship. "At least one (one) of the following" or a similar expression means any combination of these items, including any combination of single items (ones) or plural items (ones). For example, at least one (one) of a, b, or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0106] In several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units is only a logical function division, and there can be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection to each other can be through some interfaces. The indirect coupling or communication connection of devices or units can be in electrical, mechanical, or other forms.

[0107] The unit described as a separation component may or may not be physically separated, and the component shown as a unit may or may not be a physical unit, that is, it may be located in one place or may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0108] In addition, each functional unit in various embodiments of the present application may be integrated in a processing unit, may exist separately as individual physical units, or two or more units may be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0109] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions for causing an electronic device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in various embodiments of the present application. The aforementioned storage medium includes: various media that can store programs such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0110] The above is a specific description of the preferred implementation of the embodiments of the present application, but the embodiments of the present application are not limited to the above-mentioned implementation manners. Those skilled in the art can also make various equivalent deformations or substitutions without departing from the spirit of the embodiments of the present application, and these equivalent deformations or substitutions are all included within the scope defined by the claims of the embodiments of the present application.

Claims

1. A joint forecasting method for multiple sea ice elements based on causal sharing, characterized in that, The method includes: Obtaining n different sea ice element data, and dividing the n groups of different sea ice element data into s sea ice element data groups; each sea ice element data group includes 2 different sea ice element data, and any two of the n different sea ice element data form a sea ice element data group, where n and s are both positive integers greater than 2; Extracting the two-dimensional spatial field between the time series corresponding to every two different sea ice element data as the causal relationship value according to the causal information flow theory; Constructing a joint prediction model and training the joint prediction model; the joint prediction model includes s causal relationship modules and n branch networks with the same structure, where training the joint prediction model includes: Inputting the n different sea ice element data into a corresponding one of the branch networks respectively to obtain the first feature map corresponding to one sea ice element data extracted by each branch network; Extracting the causal relationship feature map corresponding to a group of sea ice element data groups by each causal relationship module according to the corresponding first feature map and the causal relationship value; Inputting the causal relationship feature maps into the two branch networks corresponding to a group of sea ice element data groups respectively, so that the branch networks generate the sea ice element prediction results corresponding to one of the sea ice element data according to the corresponding first feature map and the causal relationship feature map; Updating the network parameters of the branch networks according to the sea ice element prediction results until the trained joint prediction model is obtained; Performing joint prediction on n different target sea ice element data according to the trained joint prediction model.

2. The joint prediction method for multi-element sea ice based on causal sharing according to claim 1, characterized in that Extracting the causal relationship value between every two different sea ice element data through the following formula: ； Among them, represents a time series of sea ice element data for the time series of another sea ice element data causality value of represents determinant calculation represents the covariance matrix represents the number of variables is the variable represents cofactor of represents all and the sample covariance between the Euler forward difference approximation of represents and the sample covariance between represents the sample variance of 3. The joint prediction method for multi-element sea ice based on causal sharing according to claim 1, wherein The joint prediction model further includes a CBAM layer, and the output end of the CBAM layer is connected to the input end of each causal relationship module; Before extracting the causal relationship feature map corresponding to a group of sea ice element data groups according to the corresponding first feature map and the causal relationship value, it further includes: Inputting the n first feature maps into the CBAM layer to enhance the attention mechanism of the n first feature maps respectively through the CBAM layer to obtain n enhanced first feature maps; The extracting the causal relationship feature map corresponding to a group of sea ice element data groups according to the corresponding first feature map and the causal relationship value includes: Extracting the causal relationship feature map corresponding to a group of sea ice element data groups according to the corresponding enhanced first feature map and the causal relationship value.

4. The joint prediction method for multi-element sea ice based on causal sharing according to claim 3, wherein The causal relationship feature map of a group of sea ice element data groups includes: the first causal relationship feature map corresponding to the first sea ice element data in a group of sea ice element data groups and the second causal relationship feature map corresponding to the second sea ice element data; The extracting the causal relationship feature map corresponding to a group of sea ice element data groups according to the corresponding enhanced first feature map and the causal relationship value includes: Multiply the first enhanced first feature map corresponding to the first type of sea ice element data by the first causal relationship value to obtain a first intermediate feature map, and add the first intermediate feature map and the first enhanced first feature map to obtain the first causal relationship feature map; the first causal relationship value is the causal relationship value of the second type of sea ice element data to the first type of sea ice element data; Multiply the second enhanced first feature map corresponding to the second type of sea ice element data by the second causal relationship value to obtain a second intermediate feature map, and add the second intermediate feature map and the second enhanced first feature map to obtain the second causal relationship feature map; the second causal relationship value is the causal relationship value of the first type of sea ice element data to the second type of sea ice element data; Input the causal relationship feature maps into two branch networks corresponding to a group of sea ice element data groups respectively, including: Input the first causal relationship feature map into one of the branch networks that extracts the first first feature map; Input the second causal relationship feature map into one of the branch networks that extracts the second first feature map.

5. The joint prediction method for multi-elements of sea ice based on causal sharing according to claim 4, wherein The network structure of any of the branch networks includes an input layer, an encoder, a decoder, and an output layer connected in sequence; Both the input layer and the output layer are 1 convolutional layer; The encoder includes convolutional layers and downsampling layers, = 2 ; The first layer and the last layer of the encoder are both convolutional layers, and two connected convolutional layers are connected between every two downsampling layers; The decoder includes convolutional layers and downsampling layers; the first layer and the last layer of the decoder are both convolutional layers, and two connected convolutional layers are connected between every two of the upsampling layers; Inputting the n different types of sea ice element data into one of the corresponding branch networks respectively to obtain the first feature map of each branch network corresponding to one type of sea ice element data, including: Input each type of sea ice element data into a corresponding one of the said branch networks respectively to obtain the first feature map output by the th downsampling layer or the th upsampling layer corresponding to one of the said branch networks; is a positive integer less than or equal to ; is a positive integer less than or equal to ; Inputting the causal relationship feature maps into two branch networks corresponding to a group of sea ice element data groups respectively, including: Fuse the first feature maps output by the th downsampling layer in the two branch networks that respectively fuse the causal relationship feature map and the corresponding set of sea ice element data groups to obtain two corresponding fused feature maps, and input each fused feature map into the first convolutional layer of the corresponding branch network; the first convolutional layer is the convolutional layer after and connected to the th downsampling layer; Or, The first feature maps output by the th upsampling layer in the two branch networks that respectively fuse the causal relationship feature map and a corresponding set of sea ice element data groups are fused to obtain two corresponding fused feature maps, and each fused feature map is respectively input into the first convolutional layer of the corresponding branch network; the first convolutional layer is a convolutional layer that is after and connected to the th upsampling layer.

6. The joint prediction method for multi-element sea ice based on causal sharing according to claim 5, wherein At least one output end of a convolutional layer in the encoder is residually connected to an input end of a convolutional layer with the same feature map dimension in the decoder.

7. The joint prediction method for multi-element sea ice based on causal sharing according to claim 1, wherein The n is 3, and the n different types of sea ice element data include sea ice concentration data, sea ice velocity data, and sea ice thickness data.

8. A joint prediction system for multi-element sea ice based on causal sharing, characterized in that, The system includes: A data acquisition unit, configured to acquire n different types of sea ice element data, and divide the n groups of different sea ice element data into s sea ice element data groups; each sea ice element data group includes 2 different types of sea ice element data, and any two of the n different types of sea ice element data form a sea ice element data group, and both n and s are positive integers greater than 2; A relationship value acquisition unit, configured to extract a two-dimensional spatial field between time series corresponding to each two different types of sea ice element data as a causal relationship value according to the causal information flow theory; A model training unit, configured to construct a joint prediction model and train the joint prediction model; the joint prediction model includes s causal relationship modules and n branch networks with the same structure, wherein training the joint prediction model includes: Inputting the n different types of sea ice element data into one of the corresponding branch networks respectively to obtain the first feature map of each branch network corresponding to one type of sea ice element data; Each of the causal relationship modules extracts a causal relationship feature map of a corresponding set of sea ice element data groups according to the corresponding first feature map and the causal relationship value; The causal relationship feature maps are respectively input into two branch networks of a corresponding set of sea ice element data groups, so that the branch networks generate sea ice element prediction results corresponding to a type of the sea ice element data according to the corresponding first feature map and the causal relationship feature map; The network parameters of the branch networks are updated by gradients according to the sea ice element prediction results until a trained joint prediction model is obtained; A model application unit is configured to perform joint prediction on n different target sea ice element data according to the trained joint prediction model.

9. An electronic device, characterized in that: It includes at least one controller and a memory communicatively connected to the at least one controller; the memory stores instructions executable by the at least one controller, and the instructions are executed by the at least one controller so that the at least one controller can execute the joint prediction method for multi-elements of sea ice based on causal sharing according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-executable instructions for causing a computer to execute the joint prediction method for multi-elements of sea ice based on causal sharing according to any one of claims 1 to 7.

Images