Arctic sea ice concentration prediction method based on Swin Transform

Abstract

The invention discloses a north pole sea ice concentration prediction method based on Swin Transform, and belongs to the technical field of remote sensing data analysis. According to the method, the land mask image is generated to shield the land area, the spatial-temporal feature processing is fused, and the Swin Transform network structure is utilized to reduce the model calculation amount and enhance the perception ability of the model to the spatial-temporal features, so that the model has relatively high robustness. According to the invention, through analyzing the temporal and spatial change rule of sea ice coverage, accurate prediction and monitoring of the sea ice concentration in a period of time in the future are realized, and a reliable decision basis is provided for the fields of shipping safety, resource development, ecological environment protection, marine resource management and the like of the arctic sea area.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing data analysis, and in particular relates to a method for predicting Arctic sea ice density based on Swin Transformer. Background Art

[0002] Sea ice is a layer of ice formed on the surface of the sea, mainly composed of frozen seawater. Its formation and evolution process is complexly affected by multiple factors such as climate change, ocean circulation, and wind. Changes in Arctic sea ice cover are one of the important indicators of global climate change and have a profound impact on the stability and changes of the Earth's climate system. By monitoring and understanding the changing patterns of Arctic sea ice, we can improve our ability to predict climate change, thereby providing important data support for climate change research and the opening of summer Arctic shipping routes.

[0003] In recent years, the area and thickness of Arctic sea ice have continued to shrink. Opening up commercial shipping routes in the Arctic in summer can significantly shorten the shipping distance from Asia to Europe. Accurately predicting the daily sea ice density in the Arctic and the distribution of sea ice in the next 14 days will help ships avoid sea ice areas, reduce the risk of accidents, and improve shipping safety. In addition, by predicting the sea ice density in the next 14 days, researchers can make long-term plans and decisions.

[0004] Arctic sea ice prediction based on remote sensing data relies on a variety of technologies. The Arctic region mainly relies on satellite remote sensing data to obtain sea ice information, including coarse-resolution images provided by SMMR, SSM / I and SSMIS satellites and ERA5 reanalysis data. The key step in the prediction process is to extract and analyze the characteristics of sea ice. Image processing technology is used to denoise, detect edges and extract features from sea ice data to obtain characteristic information such as the morphology and texture of sea ice.

[0005] Machine learning and deep learning technologies play an important role in sea ice density prediction. Traditional machine learning methods such as support vector machines (SVM) and random forests are used to build statistical models, while deep learning methods such as convolutional neural networks (CNN) and recurrent neural networks (RNN) better capture the spatiotemporal features in images and improve the precision and accuracy of predictions. Data assimilation technology is a key method for integrating multi-source observation data with model prediction results. Through data assimilation, observation data from different sources can be effectively integrated, prediction errors can be reduced, and the accuracy and reliability of sea ice density predictions can be improved. The development and innovation of these technologies provide strong support for improving the accuracy and reliability of sea ice predictions.

[0006] Although most current sea ice density predictions based on numerical or statistical methods have achieved good results, these methods still have some problems. The main ones are:

[0007] (1) Complex nonlinear relationships: Sea ice data contain complex nonlinear relationships and spatiotemporal characteristics. How to adaptively extract these complex nonlinear relationships, automatically learn the features and patterns in the input data, and improve the flexibility and accuracy of the model is a difficult problem that needs to be solved urgently.

[0008] (2) Computational resource consumption: Traditional numerical or statistical models require a large amount of CPU computing resources and cannot perform lightweight predictions. Therefore, how to more effectively capture the spatiotemporal characteristics of sea ice data and reduce the consumption of computing resources by traditional models is a task that cannot be ignored. Summary of the invention

[0009] In view of the problems of insufficient spatiotemporal feature capture and excessive computing resource requirements in the existing sea ice density prediction based on remote sensing data, this paper proposes an Arctic sea ice density prediction method based on Swin Transformer to improve the performance and accuracy of sea ice density prediction tasks.

[0010] In order to achieve the above object, the specific technical scheme adopted by the present invention is as follows: A method for predicting Arctic sea ice density based on Swin Transformer includes the following steps: S1: Collect historical sea ice density observation data, perform preprocessing, extract sea ice density features, and construct training data sets, evaluation data sets, and test data sets; S2: Construct a Swin Transformer network model. The core architecture of this network model includes the Encoder module, the Decoder module, and the Swin Transformer module in the middle. Specifically, it is as follows: input data → Encoder module extracts features → Swin Transformer module processes features → Decoder module restores and reconstructs data; S3: Send the training data set to the Swin Transformer network model for training, and train the spatiotemporal features during the training process; S4: Use the evaluation dataset to verify the performance of the Swin Transformer network model and dynamically adjust the model parameters to achieve the best results at the end of training; S5: Use the trained Swin Transformer network model to test the test data set, and use qualitative and quantitative indicators to comprehensively evaluate the prediction results of the model; S6: Output the final predicted result graph.

[0011] Furthermore, the S1 is specifically: S1-1: Download the OSI-450-a dataset from the OSI-SAF official website. This dataset provides global sea ice density information from 1978 to 2020. Process the dataset according to the official documentation and use Python Request encoding to batch modify the downloaded data naming format for easy reading and processing in the next stage. S1-2: Preprocess the collected sea ice density dataset, use the netCDF4 and xarray library functions in Python to extract the temporal and spatial features in the dataset, sort the temporal features in chronological order and generate a time list to ensure that all temporal features are within the specified start and end time periods; The land data values ​​in the original dataset are all set to , encode according to this data feature, use code to process the land part in the data set, generate a land mask based on the data value of the land part, ensure that the land part data does not need to be loaded when the model is used for training in the later stage, thereby reducing the calculation amount of the model, and finally, preprocess all the land values ​​to 0, reducing the space occupied by storing land values ​​and facilitating model recognition; S1-3: Sea ice density is the fraction of ocean area covered by sea ice. It is given as a percentage, ranging from 0 to 100. Normalization is used to process all sea ice data to be distributed in Finally, the preprocessed data is written into the "npy" file and output as the data set to be used; S1-4: Divide the preliminarily processed data set into the data set used in the training phase, the data set used in the model performance evaluation phase, and the data set used in the final result testing phase of the model.

[0012] Furthermore, in S2, the data input into the Swin Transformer network model is represented as a five-dimensional feature of [B, T, C, H, W], where B represents the Batch Size of the model, T represents the image input into the model at one time, C represents the input channel dimension, and H and W represent the length and width of the image input into the model; in order to facilitate subsequent processing of the feature dimension, the feature dimension is Reshaped to [B*T, C, H, W].

[0013] Furthermore, the processing process of the Swin Transformer network model is specifically as follows: S2-1: In the Encoder stage, N serial convolutions are designed to extract the spatial feature information in the sea ice data. The calculation process of the Encoder is expressed as: ; in Indicates The output features of the Encoder module of the block, Representation layer normalization, Indicates that the convolution kernel size is The two-dimensional convolution of express Activation function; during the downsampling process, the dimensions of the data input to the model change from [56, 4, 216, 216] → [56, 64, 216, 216] → [56, 64,108, 108] → [56, 64, 108, 108] → [56, 64, 54, 54]; S2-2: Send the extracted features to A stacked and interactive Swin Transformer module is used to fuse the time dimension and the space dimension to obtain a fused feature with the shape of [B, T*C, H, W], which is processed by Swin Transformer. The calculation of the Swin Transformer module is expressed as: ; in and Respectively represent Output features of the MSA module and MLP module of the block; W-MSA and SW-MSA represent window-based multi-head self-attention using regular and shifted window partition configurations, respectively. Representation layer normalization; S2-3: In the Decoder stage, N serial deconvolution blocks are used to restore the image. The feature dimension of the data changes in the following order: [56, 64, 54, 54] → [56, 64, 108, 108] → [56, 64, 108, 108] → [56, 64, 216, 216] → [56, 4, 216, 216]. The deconvolution block includes a convolution kernel size of Deconvolution layer, a group normalization layer and a Activation Function , the specific calculation is as follows: ; in Indicates Output features of the Decoder module of the block; S2-4: Use Huber Loss to calculate the loss of the final output result, which is defined as follows: ; in, is the true value, is the predicted value, is the sample size, Represents the hyperparameters of Huber Loss; S2-5: Use AdamW optimization algorithm for parameter optimization.

[0014] Furthermore, in S3, the divided data set is sent to the network model for training, and the spatiotemporal features are trained during the training process: the training for spatiotemporal features can enable the model to learn different types of feature representations at the same time, thereby improving the model's ability to understand the data. The time feature can capture the dynamic changes of the sequence data, while the spatial feature can capture the structure and morphological information of the spatial data. This fusion training design can also effectively reduce the model's excessive dependence on a certain feature, thereby reducing the risk of overfitting, balancing the contribution of the two features to the model, and thus improving the generalization ability of the model.

[0015] Furthermore, in S4, an iterative update and early stopping strategy is used during evaluation training, including the following steps: S4-1: Use the divided evaluation data set to evaluate the performance of the model after one epoch of training. Take the result of the first training of the model as the benchmark. If the effect of subsequent training is better than this benchmark, update the model, otherwise retrain the model. S4-2: When the number of rounds in which the benchmark indicator has not been updated accounts for 10% of the total number of rounds during training, it can be determined that the model performance has reached the best, and the training is stopped. The trained model is saved and prepared for the next stage of testing.

[0016] Compared with the prior art, the advantages and beneficial effects of the present invention are: The present invention uses the land information in the data set to generate a mask image, and uses the mask image to cover the land area during training, so that the model only needs to focus on the ocean area when performing prediction tasks, thereby greatly reducing the computational complexity of the model. The present invention uses the Swin Transformer network structure to process the extracted spatiotemporal features, and uses the window-based multi-head self-attention with conventional and shifted window partition configurations to fully learn the dependencies between features, enhance the model's ability to represent spatiotemporal features and its ability to understand the trend of sea ice changes, and improve the model's prediction ability.

[0017] The present invention achieves accurate prediction and monitoring of sea ice density in the future by analyzing the temporal and spatial variation patterns of sea ice coverage, providing a reliable basis for decision-making in the fields of shipping safety, resource development, ecological environment protection, and marine resource management in the Arctic waters. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic diagram of a flowchart of an embodiment of the present invention.

[0019] Figure 2 It is a schematic diagram of a principle flow chart of an embodiment of the present invention.

[0020] Figure 3 Schematic diagram of the neural network structure described in an embodiment of the present invention.

[0021] Figure 4 Detailed structural diagram of the Swin Transformer module according to an embodiment of the present invention.

[0022] Figure 5 A schematic diagram of part of input data according to an embodiment of the present invention.

[0023] Figure 6 This is the actual sea ice density distribution used in the embodiments of the present invention.

[0024] Figure 7 This is the prediction of the sea ice density for the next 14 days according to an embodiment of the present invention.

[0025] Figure 8 This is the difference between the prediction of the sea ice density in the next 14 days by the comparison method IceNet used in the embodiments of the present invention and the actual distribution.

[0026] Fig. 9 This is the difference between the prediction of the sea ice density in the next 14 days by the comparison method SICNet used in the embodiment of the present invention and the actual distribution.

[0027] Fig.10 This is the difference between the prediction of the sea ice density in the next 14 days by the comparison method IceFormer used in the embodiments of the present invention and the actual distribution.

[0028] Fig.11 This is the difference between the prediction of the sea ice density for the next 14 days by the method of the present invention in the embodiment of the present invention and the actual distribution. DETAILED DESCRIPTION

[0029] The present invention is further explained and illustrated below through specific embodiments in conjunction with the accompanying drawings.

[0030] Embodiment 1: A method for predicting Arctic sea ice density based on Swin Transformer, such as Figure 1 , Figure 2 As shown, the specific steps include: Step 1: Collect historical sea ice concentration observation data and modify the file naming format: Step 1.1: Download the OSI-450-a dataset from the OSI-SAF official website. This dataset provides global sea ice density information from 1978 to 2020. Process the dataset according to the official documentation and use Python coding to batch modify the downloaded data naming format for easy reading and processing in the next stage. Step 2: Preprocess the collected sea ice density dataset, output the processed data into the "npy" data format, and construct the training dataset, evaluation dataset, and test dataset: Step 2.1: Preprocess the data set collected in step 1.1, use the netCDF4 and xarray library functions in Python to extract the temporal and spatial features in the data set, sort the temporal features in chronological order and generate a time list, ensuring that all temporal features are within the specified start and end time periods; Step 2.2: The land data values ​​in the original data set are all set to -32767. Encoding is performed based on this data feature. The land data in the data set is processed with code. A land mask is generated based on the land data values ​​to ensure that the land data does not need to be loaded when the model is used for training in the later stage, thereby reducing the amount of calculation of the model. Finally, all land values ​​are preprocessed to 0, which reduces the space occupied by storing land values ​​and facilitates model recognition. Step 2.3: Sea ice concentration is the fraction of ocean area covered by sea ice. It is given as a percentage ranging from 0 to 100. Normalization is used to treat all sea ice data as distributed in Finally, the preprocessed data is written into the "npy" file and output as the data set to be used; Step 2.4: Divide the preliminarily processed data set into a data set used in the training phase, a data set used in the model performance evaluation phase, and a data set used in the final result test phase. The ratio of these three data sets used in this embodiment is ; Step 3: Build a Swin Transformer network model that can extract and learn spatiotemporal sequences.

[0031] The network core architecture constructed by the present invention consists of an Encoder module, a Decoder module and a SwinTransformer module in the middle. The detailed model network structure is as follows: Figure 3As shown in the figure, it is represented as: input data → Encoder module extracts features → Swin Transformer module processes features → Decoder module restores and reconstructs data; the input data is a five-dimensional feature that can be represented as [B, T, C, H, W], B represents the batch size of the model, and the batch size is set to 4 during the experiment. If the value is set too large, it will cause excessive memory usage; T represents the picture input to the model at one time. The purpose of this embodiment is to predict sea ice in the next two weeks, so the ; C represents the input channel dimension. Before inputting the data into the model, the image is cut into 4 non-overlapping sub-images of equal size and fed into the model, so the initial channel dimension is 4; H and W represent the length and width of the image input to the model. Before inputting the features into the dual-branch convolution module, the feature dimensions are reshaped to [B*T, C, H, W] for the convenience of subsequent processing of the feature dimensions.

[0032] Step 3.1: Design N serial convolutions in the Encoder stage to extract the characteristic information in the sea ice data (take ), during the downsampling process, the dimension of the data input to the model changes as follows: [56, 4,216, 216] → [56, 64, 216, 216] → [56, 64, 108, 108] → [56, 64, 108, 108] → [56, 64, 54, 54]; Step 3.2: If Figure 4 As shown, the extracted features are fed into A stacked interactive SwinTransformer module (in the experiment, ), the time dimension and space dimension of these features are fused to obtain the fused features of shape [B, T*C, H, W], which are processed by Swin Transformer. The calculation of Swin Transformer module can be expressed as: ; in and Respectively represent The output features of the MSA module and the MLP module of the block; W-MSA and SW-MSA represent the window-based multi-head self-attention using regular and shifted window partition configurations, respectively. Normalization of the representation layer. Combining these two window mechanisms can fully mine the information of sea ice data and enhance the model's ability to perceive and understand this information; Step 3.3: In the Decoder stage, use N serial deconvolution blocks to restore the image (take ), the feature dimension of the process data changes in sequence: [56, 64, 54, 54] → [56, 64, 108, 108] → [56, 64, 108, 108] → [56, 64, 216, 216] → [56, 4, 216, 216]. This deconvolution block includes a convolution kernel size of Deconvolution layer, a group normalization layer and a Activation Function , the specific calculation is as follows: ; in Indicates Output features of the Decoder module of the block; Step 3.4: Use Huber Loss to calculate the loss of the final output result. This loss function is also called smooth mean absolute error. Compared with the mean square error (MSE), Huber Loss is more robust to outliers when the error is small because it is a weighted combination of absolute error and square error. Huber Loss has less impact on outliers in some cases, while retaining the continuity and differentiability of MSE. It is defined as follows: ; in, is the true value, is the predicted value, is the sample size, Represents the hyperparameters of Huber Loss; Step 3.5: Use the AdamW optimization algorithm to optimize parameters. Compared with the standard Adam algorithm, AdamW is more accurate in handling weight decay, can better control the regularization effect of the model, and improve the generalization performance of the model.

[0033] Step 4: Send the divided data set to the network model for training. During the training process, train the spatiotemporal features: Training on temporal and spatial features is a fusion process. By fusing temporal and spatial features, the model can learn different types of feature representations at the same time, improving the model's ability to understand data. Temporal features can capture the dynamic changes of sequence data, while spatial features can capture the structure and morphological information of spatial data. This fusion training design can also effectively reduce the model's excessive reliance on a certain feature, thereby reducing the risk of overfitting, balancing the contribution of the two features to the model, and thus improving the model's generalization ability.

[0034] Step 5: Use the evaluation data set to verify the performance of the network model and dynamically adjust the model parameters to achieve the best effect at the end of training: Step 5.1: Use the evaluation data set divided in step 2 to evaluate the performance of the model after 1 epoch training. Take the result of the first training of the model as the benchmark. If the effect of subsequent training is better than this benchmark, update the benchmark, otherwise retrain the model. Step 5.2: When the number of rounds in which the benchmark indicator has not been updated accounts for 10% of the total number of rounds during training, it can be determined that the model performance has reached the best, and training is stopped. The trained model is saved and prepared for the next stage of testing.

[0035] Step 6: Use the model trained in step 5 to test the test data set, and use qualitative and quantitative indicators to comprehensively evaluate the model's prediction results.

[0036] Step 7: Use the trained model to make predictions and output the final prediction result graph.

[0037] Embodiment 2: This embodiment further illustrates the present invention through simulation experiments: This embodiment is carried out in the hardware environment of AMD EPYC 7713, NVIDIA L40, memory 770 GB and the software environment of Ubuntu24.04.1 LTS, Python 3.12.7, Pytorch 2.5.1. The simulation experiment data of this embodiment is the OSI-450-a dataset downloaded from the OSI-SAF website. The dataset includes coarse-resolution images provided by SMMR, SSM / I and SSMIS satellites and ERA5 reanalysis data, providing global sea ice density information from 1978 to 2020. The size of each image in the dataset is Pixels, such as Figure 5As shown in the figure, a schematic diagram of Arctic sea ice density under representative extreme weather conditions is shown, namely April 1, 2020, June 1, 2020, August 1, 2020 and September 1, 2020. By selecting these representative dates, the importance and application value of sea ice density prediction can be fully demonstrated. At the same time, it can also help users better understand the complexity and changing laws of the sea ice system, and demonstrate the important impact of sea ice density on the environment and human activities.

[0038] The method of the present invention is compared with the existing advanced sea ice concentration prediction methods. The IceNet method used in the comparative test is proposed in the article "Seasonal Arctic sea ice forecasting with probabilistic deep learning", the SICNet method is proposed in the article "A Data-Driven Deep Learning Model for Weekly Sea Ice Concentration Prediction of the Pan-Arctic During the Melting Season", and the IceFormer method is proposed in the article "A Spatiotemporal Multiscale Deep Learning Model for Subseasonal Prediction of Arctic Sea Ice".

[0039] like Figure 6 As shown in the figure, for the OSI-450-a dataset, the method of the present invention can effectively predict the distribution of Arctic sea ice density in the future. Figure 7 The actual sea ice density distribution over these 14 days is shown; Figure 8 , Fig. 9 and Fig.10 The differences between the prediction results of IceNet, SICNet and IceFormer models and the actual distribution of sea ice density are shown respectively. Fig.11 The difference between the sea ice density distribution predicted by the method of the present invention for the next 14 days and the actual sea ice density is demonstrated. By comparing the difference results, it can be seen that the prediction effect obtained by using the method of the present invention is very close to the actual situation, which once again verifies the superiority of the method of the present invention.

[0040] The evaluation indicators used in the quantitative analysis of the model performance in the present invention are mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE), Nash-Sutcliffe efficiency coefficient (NSE) and peak signal-to-noise ratio (PSNR), which are calculated as follows: ; ; ; in, is the observed value, is the corresponding predicted value, Represents the number of samples. The smaller the MSE value, the smaller the difference between the predicted value and the observed value, and the better the model fitting effect. Compared with MSE, RMSE is more intuitive. The smaller the MAE, the smaller the difference between the predicted value and the observed value, and the better the model fitting effect.

[0041] ; in, It is the average of the observed values. The value range of NSE is from negative infinity to 1. The closer the value is to 1, the better the model performance.

[0042] ; in, is the maximum possible value of the data, where , MSE is the mean square error, and the higher the PSNR value, the higher the similarity between the prediction result and the original data.

[0043] Tables 1, 2, 3, 4 and 5 respectively show the comparison results of MSE, RMSE, MAE, NSE and PSNR of these model methods in the five years from 2016 to 2020. It can be seen from the above tables that the method of the present invention is superior to other model methods in these five evaluation indicators, and has achieved good prediction results during these five years, which shows that the method of the present invention can more accurately predict the density of Arctic sea ice in the future.

[0044] Table 1 Comparison of performance indicators of prediction results of OSI-450-a dataset in 2016 .

[0045] Table 2 Comparison of performance indicators of prediction results of OSI-450-a dataset in 2017 .

[0046] Table 3 Comparison of performance indicators of prediction results of OSI-450-a dataset in 2018 .

[0047] Table 4 Comparison of performance indicators of prediction results of OSI-450-a dataset in 2019 .

[0048] Table 5 Comparison of performance indicators of prediction results of OSI-450-a dataset in 2020 .

[0049] The method for predicting Arctic sea ice density from remote sensing images based on the Swin Transformer framework provided by the present invention is mainly used to predict the daily distribution of Arctic sea ice. The article predicts the Arctic sea ice density for the next two weeks. However, the method of the present invention is also applicable to forecasts for the next few days or even weeks, and is also applicable to predicting the monthly distribution of Arctic sea ice density, and the beneficial effects achieved are similar.

[0050] The above description is only a preferred embodiment of the present invention and does not limit the present invention in other forms. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the protection scope of the technical solution of the present invention.

Claims

1. A method for predicting Arctic sea ice density based on Swin Transformer, characterized in that: The following steps are involved: S1: Collect historical sea ice density observation data, perform preprocessing, extract sea ice density features, and construct training data sets, evaluation data sets, and test data sets; S2: Construct a Swin Transformer network model. The core architecture of this network model includes the Encoder module, the Decoder module, and the Swin Transformer module in the middle. Specifically, it is as follows: input data → Encoder module extracts features → Swin Transformer module processes features → Decoder module restores and reconstructs data; S3: Send the training data set to the Swin Transformer network model for training, and train the spatiotemporal features during the training process; S4: Use the evaluation dataset to verify the performance of the Swin Transformer network model and dynamically adjust the model parameters to achieve the best results at the end of training; S5: Use the trained Swin Transformer network model to test the test data set, and use qualitative and quantitative indicators to comprehensively evaluate the prediction results of the model; S6: Output the final predicted result graph.

2. The method for predicting Arctic sea ice density according to claim 1, characterized in that: The S1 is specifically: S1-1: Collect data sets and modify the naming format of downloaded data; S1-2: Extract the temporal and spatial features in the data set, sort the temporal features in chronological order and generate a time list, ensuring that all temporal features are within the specified start and end time periods; S1-3: Sea ice density is normalized to distribute all sea ice data in Finally, the preprocessed data is written into the "npy" file and output as the data set to be used; S1-4: Divide the preliminarily processed data set into the data set used in the training phase, the data set used in the model performance evaluation phase, and the data set used in the final result testing phase of the model.

3. The method for predicting Arctic sea ice density according to claim 1, characterized in that: In S2, the data input into the SwinTransformer network model is represented as a five-dimensional feature of [B, T, C, H, W], where B represents the batch size of the model, T represents the image input into the model at one time, C represents the input channel dimension, and H and W represent the length and width of the image input into the model; in order to facilitate subsequent processing of the feature dimension, the feature dimension is reshaped to [B*T, C, H, W].

4. The method for predicting Arctic sea ice density according to claim 1, characterized in that: The processing process of the Swin Transformer network model is specifically as follows: S2-1: In the Encoder stage, N serial convolutions are designed to extract the spatial feature information in the sea ice data. The calculation process of the Encoder is expressed as: ； in Indicates The output features of the Encoder module of the block, Representation layer normalization, Indicates that the convolution kernel size is The two-dimensional convolution of express Activation function; S2-2: Send the extracted features to The stacked Swin Transformer modules are used to fuse the time dimension and the space dimension to obtain a fused feature with the shape of [B, T*C, H, W], which is processed by Swin Transformer. The calculation of the Swin Transformer module is expressed as: ； in and Respectively represent Output features of the MSA module and MLP module of the block; W-MSA and SW-MSA represent window-based multi-head self-attention using regular and shifted window partition configurations, respectively. Representation layer normalization; S2-3: In the Decoder stage, N serial deconvolution blocks are used to restore the image. The deconvolution block includes a convolution kernel size of Deconvolution layer, a group normalization layer and a Activation Function , the specific calculation is as follows: ； in Indicates Output features of the Decoder module of the block; S2-4: Use Huber Loss to calculate the loss of the final output result, which is defined as follows: ； in, is the true value, is the predicted value, is the sample size, Represents the hyperparameters of Huber Loss; S2-5: Use AdamW optimization algorithm for parameter optimization.

5. The method for predicting Arctic sea ice density according to claim 1, wherein: In S4, an iterative update and early stopping strategy is used during evaluation training, including the following steps: S4-1: Use the divided evaluation data set to evaluate the performance of the model after one epoch of training. Take the result of the first training of the model as the benchmark. If the effect of subsequent training is better than this benchmark, update the model, otherwise retrain the model. S4-2: When the number of rounds in which the benchmark indicator has not been updated accounts for 10% of the total number of rounds during training, it can be determined that the model performance has reached the best, and training is stopped. The trained model is saved and prepared for the next stage of testing.

Images