Intelligent downscaling method for polar region sea ice concentration based on deep self-attention residual U-Net

Abstract

The invention belongs to the technical field of polar region marine environment element downscaling, and particularly relates to a polar region sea ice concentration intelligent downscaling method based on a deep self-attention residual U-Net, and the method comprises the steps: collecting and preprocessing the historical statistical reanalyzed sea surface temperature and sea ice data; 2, constructing a data set of network training; step 3, establishing a DeepSARU-Net network architecture, and setting parameters of an optimizer; step 4, selecting an evaluation standard, judging a downscaling effect of the DeepSARU-Net network downscaling method, optimizing DeepSARU-Net network architecture parameters, and outputting a trained DeepSARU-Net network architecture; and 5, performing normalization processing on to-be-processed sea surface temperature and sea ice data of the target sea area, performing processing by using the trained DeepSARU-Net network architecture, and outputting a prediction result of the sea ice concentration of the target sea area. According to the method, the low-resolution SIC data can be converted into high-resolution data with richer spatial details and higher physical consistency.

Description

Technical Field

[0001] The present invention belongs to the technical field of polar ocean environmental element downscaling, and specifically relates to an intelligent downscaling method for polar sea ice density based on deep self-attention residual U-Net. Background Art

[0002] As a key component of the climate system, sea ice has a significant impact on the ecosystem dynamics and ocean circulation patterns in the polar regions, and its effects have far-reaching impacts on global temperature fluctuations and sea level changes. Against this macro background, it has become increasingly important and urgent to achieve high-resolution predictions of polar marine environmental factors. Sea ice concentration (SIC) is an important indicator for measuring the status of polar sea ice coverage and is of great significance to polar shipping, climate change research, marine ecosystem monitoring and other fields. However, current sea ice concentration observation data generally have problems such as low spatial resolution, spatiotemporal discontinuity, and limited accuracy, making it difficult to meet the needs of refined applications. Therefore, the development of high-precision sea ice concentration intelligent downscaling (super-resolution) methods is crucial to improving sea ice monitoring accuracy, optimizing polar shipping route planning, and improving marine environment simulation capabilities.

[0003] Current methods for downscaling sea ice concentration are primarily divided into numerical modeling methods based on ocean dynamic equations and mathematical statistics methods based on probabilistic statistical models. The former relies on coupled sea ice-ocean models (such as CICE, HYCOM, and PIMOS) to simulate sea ice evolution by solving sea ice dynamic and thermodynamic equations. These methods can accurately characterize the large-scale distribution characteristics of sea ice and offer advantages in terms of physical consistency. However, because numerical models are often limited by initial conditions, parameterization schemes, and computational resources, their simulation results often suffer from significant uncertainty and are limited in their ability to accurately depict sea ice details. In particular, the accuracy of traditional numerical models for the sea ice edge, the land-sea interface, and complex sea ice dynamic processes (such as ridge formation, ice crack propagation, and the density distribution within the sea ice) remains insufficient for practical applications.

[0004] Mathematical statistics methods based on probability statistics mainly rely on historical observational data and downscale predictions of sea ice density by constructing empirical models or machine learning methods. For example, multiple regression analysis, time series models (such as ARIMA), and spatial statistical methods based on Kriging interpolation have been widely used to complete the spatial space and improve the resolution of sea ice data. Such methods can, to a certain extent, explore the patterns in the observational data and reduce computational costs. However, due to their reliance on historical data training, their ability to model nonlinear and non-stationary sea ice changes is limited. In the case of extreme climate events or large changes in data distribution, their generalization ability is insufficient, making it difficult to adapt to the prediction needs of different regions and spatiotemporal scales.

[0005] In recent years, with the development of deep learning technology, data-driven deep learning methods have gradually become an important direction in sea ice concentration downscaling research. Compared with traditional methods, deep learning can automatically learn the complex mapping relationships of sea ice characteristics and combine multi-source data for information fusion, thereby demonstrating superior performance in super-resolution reconstruction of sea ice concentration. In particular, the introduction of deep learning techniques such as self-attention mechanism, residual learning, and multi-scale feature extraction has enabled models to achieve breakthrough progress in capturing the dynamic evolution of sea ice, improving the accuracy of super-resolution predictions, and enhancing cross-regional generalization capabilities. Therefore, developing a deep learning-based intelligent downscaling method has become a key trend in current sea ice concentration super-resolution research.

[0006] This paper aims to meet the needs of ensuring polar marine environmental factors for tasks such as Arctic navigation, Arctic waterway risk assessment and Arctic environmental monitoring. It designs an intelligent downscaling method for sea ice density that integrates a self-attention mechanism, a residual network and a U-Net architecture, which improves the downscaling accuracy of marine environmental factors and the expression of detailed scales. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent downscaling method for polar sea ice density based on deep self-attention residual U-Net, which can improve the downscaling accuracy of marine environmental elements and the expression of detail scale.

[0008] The technical solutions adopted by the present invention are as follows:

[0009] An intelligent downscaling method for polar sea ice density based on deep self-attention residual U-Net includes the following steps:

[0010] Step 1: Collect historical statistical reanalysis sea surface temperature and sea ice data and perform normalization preprocessing operations;

[0011] In step 1, the ocean data information based on the reanalysis of historical statistics of the target sea area includes sea surface temperature and sea ice data. The influence of the long-term average climate state on the sea ice is removed to obtain the outlier data of the sea ice, thereby highlighting the short-term fluctuations and seasonal changes of the sea ice; and the data is normalized and standardized to enable faster convergence and more stable prediction effects in the learning of the neural network.

[0012] Step 2: Construct a data set for network training and divide it into training set and test set;

[0013] Preferably, in step 2, a data set for network training is constructed, and the data is divided into a training set and a test set; after the data is preprocessed, a data set suitable for network training and verification is constructed; most of the data is divided into a training set according to a specific ratio for iterative training of the network; and a smaller part is used as a test set to evaluate the fitting effect of the network model on non-training set data in real time, and adjust the network parameters accordingly; the division of the data set not only helps in network training and verification, but also can be used as control data for comparative experiments, for example, it is convenient to compare the downscaling effects of traditional spatial downscaling methods and new methods on the same data set.

[0014] Step 3: Build the DeepSARU-Net network architecture and set the optimizer parameters;

[0015] Preferably, in the step 3, the DeepSARU-Net network architecture is constructed, and when the optimizer parameters are set, the main network of DeepSARU-Net is based on the encoding-decoding architecture of U-Net, which includes skip connections; during the downsampling operation, the stride convolution operation is used to replace the original maximum pooling operation of U-Net; at the same time, the square convolution kernel is decomposed into two mutually perpendicular rectangular convolution kernels to increase the depth of the network and capture more features at different resolution stages; through these operations, the model's ability to capture image features can be improved, which is conducive to reducing the size of the input signal in subsequent network layers, while expanding the feature receptive field range, thereby improving the performance and effect of the model.

[0016] Since the attention mechanism has become an important technology widely adopted in the field of deep learning, its core purpose is to improve the efficiency and accuracy of neural networks in processing time series data. By introducing this technology, the DeepSARU-Net network architecture introduces the attention mechanism. When performing time series related tasks, the neural network can more effectively focus on the key information at different time nodes, thereby optimizing the task execution effect. The DeepSARU-Net network architecture is designed with multiple resolution stages. In the jump connection between the first and final resolution stages, the present invention innovatively introduces the self-attention mechanism. When constructing the mathematical framework of the attention mechanism, three key vectors are first introduced: query tensor (Query, Q), key tensor (Key, K) and value tensor (Value, V). Among them, the query tensor Q represents the information currently being paid attention to and is the input of the attention mechanism. The key tensor K is used to measure the degree of correlation between the query tensor Q and the rest of the information. The value tensor V carries the actual information associated with each query tensor Q. The mathematical expression of the attention mechanism is:

[0017]

[0018] In the above formula, d k Represents the dimension of K, and each row of the attention score matrix corresponds to the attention score calculated between the i-th row of the query vector Q and each column of the key vector K.

[0019] To ensure the rationality and effectiveness of weight distribution, the attention scores are usually processed with a softmax operation to map the score values in each row to the interval [0, 1] and ensure that the sum of all elements in each row is strictly equal to 1; DeepSARU-Net uses a convolutional neural network based on U-Net as its main architecture and incorporates an attention mechanism to enhance the ability to capture temporal scale features and improve the expression accuracy of spatial scale features.

[0020] In the convolutional neural network part, residual connections are introduced to optimize the gradient propagation path; thereby achieving fine downscaling processing of sea ice density; the network is named DeepSARU-Net; by adopting the skip connection structure, the present invention can directly connect the data at the same resolution in the encoder and decoder stages; this design enables the network to better understand and utilize the multi-scale data features involved in the encoding process, thereby providing a solid foundation for in-depth analysis of the physical evolution of sea ice.

[0021] Step 4: Select evaluation criteria to judge the downscaling effect of the DeepSARU-Net network downscaling method, optimize the DeepSARU-Net network architecture parameters, and output the trained DeepSARU-Net network architecture;

[0022] Preferably, in step 4, the root mean square error (RMSE), correlation coefficient (r), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are selected as evaluation criteria to evaluate the effect of DeepSARU-Net network downscaling, and the DeepSARU-Net network architecture parameters are reversely optimized according to the downscaling effect.

[0023] Step 5: Normalize the sea surface temperature and sea ice data to be processed in the target sea area, and use the trained DeepSARU-Net network architecture to process them and output the sea ice density prediction results of the target sea area.

[0024] The technical effects achieved by the present invention are:

[0025] This paper aims to address the urgent need for high-precision sea ice information in key tasks such as Arctic navigation, Arctic shipping risk assessment, and Arctic environmental monitoring. It proposes a sea ice density intelligent downscaling method (DeepSARU-Net) that integrates a self-attention mechanism, a residual network, and a U-Net architecture to improve the spatial resolution and detail depiction of sea ice density data. The self-attention mechanism is used to model the long-range dependencies in the spatiotemporal variations of sea ice, effectively capturing the sea ice distribution characteristics driven by large-scale environmental factors. A residual network is combined with the model to optimize deep information transfer, improving model stability and training efficiency. The U-Net architecture's encoder-decoder structure is also used to enhance the ability to extract sea ice features at different scales, ensuring that the model retains key details during the downscaling and reconstruction process. Through this innovative fusion architecture, the present invention demonstrates superior performance in the spatial super-resolution reconstruction of spatiotemporal series of sea ice density. It can convert low-resolution SIC data into high-resolution data with richer spatial details and higher physical consistency, providing more accurate data information support for polar marine environmental monitoring and early warning, and has strong academic significance and engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a flow chart of an intelligent downscaling method for polar sea ice density based on deep self-attention residual U-Net of the present invention;

[0027] Figure 2 This is a schematic diagram of the selection of the experimental sea area of the present invention;

[0028] Figure 3 Schematic diagram of the DeepSARU-Net network structure of the present invention;

[0029] Figure 4This is the actual effect diagram of the DeepSARU-Net architecture downscaling of the present invention;

[0030] Figure 5 This is a flow chart of the DeepSARU-Net spatial downscaling training model of the present invention. DETAILED DESCRIPTION

[0031] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is described in detail below with reference to the following examples. It should be understood that the following text is only used to describe one or more specific embodiments of the present invention and does not strictly limit the scope of protection of the present invention.

[0032] like Figure 1-Figure 5 As shown in the figure, an intelligent downscaling method for polar sea ice density based on deep self-attention residual U-Net includes the following steps:

[0033] Step 1: Collect historical statistical reanalysis sea surface temperature and sea ice data and perform normalization preprocessing operations;

[0034] In step 1, the ocean data information based on the reanalysis of historical statistics of the target sea area includes sea surface temperature and sea ice data. The influence of the long-term average climate state on the sea ice is removed to obtain the outlier data of the sea ice, thereby highlighting the short-term fluctuations and seasonal changes of the sea ice; and the data is normalized and standardized to enable faster convergence and more stable prediction effects in the learning of the neural network.

[0035] In this paper, the main area of the validation experiment was precisely defined as the Arctic region (longitude: -180°E to 180°E, latitude: 60°N to 90°N), which includes key areas for Arctic sea ice dynamics and observational research. This region is subject to a series of complex physical mechanisms, including but not limited to short-term meteorological fluctuations and long-term climate trends. In addition, frequent commercial activities and scientific expeditions also have a significant impact on it, resulting in highly nonlinear characteristics in the evolution of sea ice on multiple time scales.

[0036] This paper focuses on the Chukchi Sea region of the Arctic Ocean (66°N to 74°N, 164°W to 180°E), and uses the OSTIA sea surface temperature and sea ice analysis data provided by CMEMS to study high-resolution sea ice density in order to reveal sea ice dynamics, assess Arctic navigation potential, and provide a scientific basis for global climate monitoring. The selected sea areas are as follows: Figure 2 shown.

[0037] This dataset has a spatial resolution of 1 / 12° × 1 / 12° and covers a continuous 20-year record from 2001 to 2020. In addition to sea ice concentration (SIC) data, this comprehensive dataset also includes important parameters such as sea ice thickness (SIT) and sea surface temperature (SST) for scientific research purposes. In this study, all data used were preprocessed to a spatial resolution of 1 / 2° × 1 / 2° to achieve a consistent analysis scale. Monthly sea ice thickness (SIT) remote sensing data are provided by the Copernicus Climate Change Service (C3S) using the RA-2 altimeter on the ENVISAT satellite and the SIRAL altimeter on the CryoSat-2 satellite. This dataset has a spatial resolution of 25 km × 25 km and covers a 20-year monthly record from 2001 to 2020.

[0038] Step 2: Construct a data set for network training and divide it into training set and test set;

[0039] Preferably, in step 2, a data set for network training is constructed, and the data is divided into a training set and a test set; after the data is preprocessed, a data set suitable for network training and verification is constructed; most of the data is divided into a training set according to a specific ratio for iterative training of the network; and a smaller part is used as a test set to evaluate the fitting effect of the network model on non-training set data in real time, and adjust the network parameters accordingly; the division of the data set not only helps in network training and verification, but also can be used as control data for comparative experiments, for example, it is convenient to compare the downscaling effects of traditional spatial downscaling methods and new methods on the same data set.

[0040] In this paper, the dataset was rigorously partitioned according to the different operational stages of the SICformer model built using this data. Specifically, reanalysis and remote sensing data from 2001 to 2016 were used as the data source for the pre-training and fine-tuning phases of the SICformer model; remote sensing data from 2017 to 2020 were reserved as an independent validation set for the model testing phase. This meticulous data partitioning strategy ensured that future sample information was strictly prevented from being leaked during model training, thereby ensuring the fairness and scientific nature of the model evaluation process and the reliability of the evaluation results.

[0041] Since the U-Net architecture adopts a multi-level symmetrical design, the design deploys multi-layer convolutional neural networks (CNNs) at each level of encoding and decoding, aiming to improve the accuracy of super-resolution reconstruction by maintaining the same resolution. However, this approach may induce gradient vanishing or gradient explosion phenomena in the multi-level configuration of the network. To meet this challenge, the present invention integrates the residual network mechanism on the basis of the original architecture. In the residual network, each residual unit contains a main path and a shortcut connection: the main path is responsible for learning the residual features, while the shortcut connection directly passes the input signal to the output without additional transformation. During the residual learning process, the shortcut connection performs an identity mapping operation, realizing direct information flow from shallow to deep layers. This learning strategy does not require the introduction of additional parameters, thereby avoiding the increase in network computational complexity and promoting the improvement of computational efficiency.

[0042] like Figure 5 As shown, step 3: build the DeepSARU-Net network architecture and set the optimizer parameters;

[0043] Preferably, in the step 3, the DeepSARU-Net network architecture is constructed, and when the optimizer parameters are set, the main network of DeepSARU-Net is based on the encoding-decoding architecture of U-Net, which includes skip connections; during the downsampling operation, the stride convolution operation is used to replace the original maximum pooling operation of U-Net; at the same time, the square convolution kernel is decomposed into two mutually perpendicular rectangular convolution kernels to increase the depth of the network and capture more features at different resolution stages; through these operations, the model's ability to capture image features can be improved, which is conducive to reducing the size of the input signal in subsequent network layers, while expanding the feature receptive field range, thereby improving the performance and effect of the model.

[0044] like Figure 5As shown in the figure, since the attention mechanism has become an important technology widely adopted in the field of deep learning, its core purpose is to improve the efficiency and accuracy of neural networks in processing time series data. By introducing this technology, the DeepSARU-Net network architecture introduces the attention mechanism. When performing time series related tasks, the neural network can more effectively focus on the key information at different time nodes, thereby optimizing the task execution effect. The DeepSARU-Net network architecture designs multiple resolution stages. In the jump connection between the first and final resolution stages, the present invention innovatively introduces the self-attention mechanism. When constructing the mathematical framework of the attention mechanism, three key vectors are first introduced: query tensor (Query, Q), key tensor (Key, K) and value tensor (Value, V). Among them, the query tensor Q represents the information currently being paid attention to and is the input of the attention mechanism. The key tensor K is used to measure the degree of correlation between the query tensor Q and the rest of the information. The value tensor V carries the actual information associated with each query tensor Q. Among them, the mathematical expression of the attention mechanism is:

[0045]

[0046] In the above formula, d k Represents the dimension of K, and each row of the attention score matrix corresponds to the attention score calculated between the i-th row of the query vector Q and each column of the key vector K.

[0047] like Figure 5 As shown in the figure, to ensure the rationality and effectiveness of weight distribution, the attention scores are usually processed with a softmax operation to map the score values in each row to the interval [0,1] and ensure that the sum of all elements in each row is strictly equal to 1; DeepSARU-Net uses a convolutional neural network based on U-Net as its main architecture and incorporates an attention mechanism to enhance the ability to capture temporal scale features and improve the expression accuracy of spatial scale features.

[0048] In the convolutional neural network part, residual connections are introduced to optimize the gradient propagation path; thereby achieving fine downscaling processing of sea ice density; the network is named DeepSARU-Net; by adopting the skip connection structure, the present invention can directly connect the data at the same resolution in the encoder and decoder stages; this design enables the network to better understand and utilize the multi-scale data features involved in the encoding process, thereby providing a solid foundation for in-depth analysis of the physical evolution of sea ice.

[0049] In the present invention, the main network design of DeepSARU-Net draws on the encoding-decoding architecture of U-Net and incorporates a skip connection mechanism. When performing downsampling processing, we use a strided convolution operation to replace the maximum pooling operation originally used by U-Net. In addition, the square convolution kernel is decomposed and transformed into two mutually perpendicular rectangular convolution kernels to deepen the network layer and capture more feature information at different resolution stages. These improvements effectively enhance the model's ability to extract image features, help to gradually reduce the size of the input signal in subsequent network layers, and expand the receptive field range of the features, thereby optimizing the performance and performance of the model. Through the above technical means, we can effectively capture the detailed features of different dimensions in the sequence data, hoping to improve the accuracy of the prediction of the internal density and dynamic distribution of sea ice.

[0050] In the present invention, the DeepSARU-Net network architecture is designed with multiple resolution stages. In the jump connection between the first and final resolution stages, the present invention innovatively introduces a self-attention mechanism, which significantly enhances the sensitivity and accuracy of the network in the feature extraction process. The entire processing flow begins with in-depth channel feature extraction of input data of size 16×32. Through this step, 16 critical low-resolution feature channels are extracted from the original data, laying a solid foundation for subsequent processing. The low-resolution feature data is finely processed using the advanced upsampling technology of bicubic interpolation to ensure that its size perfectly matches the target high-resolution image (80×160). The upsampled high-resolution features are then carefully input into the core architecture of DeepSARU-Net. The network encoding stage uses continuous convolution operations to gradually downsample the features. This process helps to extract deeper feature information and provide rich feature representations for the subsequent decoding stage. In the decoding stage, the deconvolution operation is used to gradually restore and significantly improve the resolution of the image to achieve downscaled prediction of sea ice density. The specific structure of the network is as follows Figure 3 As shown:

[0051] Step 4: Select evaluation criteria to judge the downscaling effect of the DeepSARU-Net network downscaling method, optimize the DeepSARU-Net network architecture parameters, and output the trained DeepSARU-Net network architecture;

[0052] Preferably, in step 4, the root mean square error (RMSE), correlation coefficient (r), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are selected as evaluation criteria to evaluate the effect of DeepSARU-Net network downscaling, and the DeepSARU-Net network architecture parameters are reversely optimized according to the downscaling effect.

[0053] In this paper, RMSE measures the sample standard deviation between the downscaling result and the true value. It is more sensitive to outliers and large errors and provides an intuitive measure of the accuracy of the model. Its formula is as follows:

[0054]

[0055] Among them, y i represents the true value of HR, x i represents the SR prediction value, i is the index of the grid, and n represents the total number of grids.

[0056] In this invention, the correlation coefficient is used to evaluate the linear correlation between the downscaling result and the true value, reflecting the consistency of the change trend between the two. Its formula is as follows:

[0057]

[0058] in, and Represent the mean of x and y, respectively. As can be seen, an r value close to 1 indicates a strong positive correlation between the predicted value and the true value, with the two changing in the same direction; a value close to -1 indicates a strong negative correlation between the two, with the two changing in opposite directions; and a value close to 0 indicates no linear relationship between the two.

[0059] In this invention, PSNR is an important indicator for measuring image quality and is often used to evaluate the quality difference between the reconstructed image and the original image. It reflects the degree of information retention during the image compression and reconstruction process. The calculation formula of PSNR is:

[0060]

[0061] Where n represents the number of bits per pixel, usually an 8-bit color image. The maximum color intensity value of a pixel is 2 n -1, for 8-bit representation, this value is equal to 255.

[0062] In this paper, the structural similarity index is used to measure the degree of similarity between the spatial structures of two images and is often used to evaluate the structural similarity between the reconstructed image and the original image. Compared with PSNR, SSIM is more consistent with the human eye's judgment of image quality. Its calculation formula is as follows:

[0063]

[0064] Among them, L(x,y), C(x,y) and S(x,y) represent brightness, contrast and structure terms respectively, μ x and μ y are the means of x and y, σ x and σ y are the variances of x and y, σxy is the covariance of x and y, C1, C2, and C3 are constants used to avoid division by zero errors. SSIM comprehensively measures the spatial structural similarity between two images from three key aspects by multiplying these three terms.

[0065] First, to optimize experimental consistency and comparability, this paper adopted a data batching strategy. Specifically, each data batch contained between 40 and 60 samples, a range chosen based on memory availability. This ensured that all experiments were performed on datasets of comparable size, enhancing the reliability and comparability of the experimental results. Furthermore, limiting the number of input and output channels to a single channel helped us focus on the characteristics of a single day's data, reducing experimental variation and enabling more accurate evaluation of model performance. When calculating the number of trainable parameters, this paper fully considered the complexity of the neural network architecture. As the number of network layers increases, computational complexity also increases, significantly impacting model training efficiency and performance. Therefore, when designing the experiments, we considered not only the downscaled prediction itself but also the entire training cycle, including loss function calculation and parameter optimization. This approach aims to provide a more comprehensive understanding of the overall performance of model training, enabling more scientific evaluation and optimization. The computational time of the downscaled prediction process was measured using the average prediction time on the validation set. This metric more accurately reflects the model's performance in real-world applications, particularly the time required for a single prediction step. This allows us to more intuitively understand the model's computational efficiency, providing strong support for subsequent model optimization and application.

[0066] This paper is dedicated to exploring the performance of the DeepSARU-Net architecture in the task of super-resolution prediction of sea ice density. In order to comprehensively evaluate its performance, this paper selected the classic U-Net architecture and the AttentionU-Net with the introduction of the attention mechanism as benchmark comparison methods. The experimental design follows a unified standard, that is, using one day's data as input to predict the sea ice density of different output days. In order to scientifically and objectively measure the prediction accuracy and image reconstruction quality of each model, we adopted statistical indicators such as the root mean square error (RMSE), correlation coefficient (r), peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). The results are shown in Table 1:

[0067] Table 1 Performance evaluation results of different models

[0068]

[0069]

[0070] Standard U-Net, AttentionU-Net and DeepSARU-Net are all deep neural network models with complex structures and excellent performance. In terms of performance on the validation set, the root mean square error (RMSE) of the standard U-Net is 0.0742, which is slightly inferior to the other two networks. AttentionU-Net has achieved performance improvement by cleverly integrating the attention module into the jump connection, and its RMSE has been reduced to 0.0693, which is about 6.6% less than the standard U-Net. This improvement can be attributed to the fact that AttentionU-Net uses more trainable parameters, and the introduction of the attention module effectively enhances the model's ability to capture key information, thereby improving overall performance. The DeepSARU-Net model proposed in the present invention optimizes the fusion method of the attention module and the residual architecture by redesigning the structure of the encoder and decoder, thereby reducing the number of trainable parameters while simplifying the network structure and effectively avoiding the risk of overfitting. Experimental results show that DeepSARU-Net demonstrates superior performance in addressing the complex problem of sea ice density super-resolution, significantly outperforming other methods across all evaluation metrics. Its RMSE is 0.0536, r is 0.9934, PSNR is 30.28dB, and SSIM is 0.9701. Compared to the standard U-Net, this method reduces RMSE by 27.76%, improves r by 1.96%, increases PSNR by 9.8%, and improves SSIM by 1.46%.

[0071] like Figure 4 As shown in (A), Figure 4 (A) is the prediction result of the target sea area on June 2, 2022. Figure 4 As shown in (B), Figure 4 (B) shows the prediction results for the target sea area on December 15, 2022. It can be seen that U-Net and AttentionU-Net demonstrate superior prediction generalization performance, with AttentionU-Net's prediction error slightly lower than that of U-Net in key areas. Notably, the DeepSARU-Net model performs particularly well overall, with the lowest residual error. This result demonstrates that the DeepSARU-Net architecture may offer unique advantages in capturing the complex and changing spatiotemporal dynamics of sea ice.

[0072] Step 5: Normalize the sea surface temperature and sea ice data to be processed in the target sea area, and use the trained DeepSARU-Net network architecture to process them and output the sea ice density prediction results of the target sea area.

[0073] This paper aims to address the urgent need for high-precision sea ice information in key tasks such as Arctic navigation, Arctic shipping risk assessment, and Arctic environmental monitoring. It proposes a sea ice density intelligent downscaling method (DeepSARU-Net) that integrates a self-attention mechanism, a residual network, and a U-Net architecture to improve the spatial resolution and detail depiction of sea ice density data. The self-attention mechanism is used to model the long-range dependencies in the spatiotemporal variations of sea ice, effectively capturing the sea ice distribution characteristics driven by large-scale environmental factors. A residual network is combined with the model to optimize deep information transfer, improving model stability and training efficiency. The U-Net architecture's encoder-decoder structure is also used to enhance the ability to extract sea ice features at different scales, ensuring that the model retains key details during the downscaling and reconstruction process. Through this innovative fusion architecture, the present invention demonstrates superior performance in the spatial super-resolution reconstruction of spatiotemporal series of sea ice density. It can convert low-resolution SIC data into high-resolution data with richer spatial details and higher physical consistency, providing more accurate data information support for polar marine environmental monitoring and early warning, and has strong academic significance and engineering application value.

[0074] The foregoing is merely a preferred embodiment of the present invention. It should be noted that those skilled in the art may make various improvements and modifications without departing from the principles of the present invention, and such improvements and modifications are also within the scope of protection of the present invention. Structures, devices, and operating methods not specifically described or explained herein shall, unless otherwise specified or limited, be implemented in accordance with conventional means in the art.

Claims

1. An intelligent downscaling method for polar sea ice density based on deep self-attention residual U-Net, characterized by: The following steps are involved: Step 1: Collect historical statistical reanalysis sea surface temperature and sea ice data and perform normalization preprocessing operations; Step 2: Construct a data set for network training and divide it into training set and test set; Step 3: Build the DeepSARU-Net network architecture and set the optimizer parameters; Step 4: Select evaluation criteria to judge the downscaling effect of the DeepSARU-Net network downscaling method, optimize the DeepSARU-Net network architecture parameters, and output the trained DeepSARU-Net network architecture; Step 5: Normalize the sea surface temperature and sea ice data to be processed in the target sea area, and use the trained DeepSARU-Net network architecture to process them and output the sea ice density prediction results of the target sea area.

2. The method for intelligent downscaling of polar sea ice density based on deep self-attention residual U-Net according to claim 1 is characterized by: In step 1, the ocean data information based on the historical statistical reanalysis of the target sea area includes sea surface temperature and sea ice data, and the influence of the long-term average climate state on the sea ice is removed to obtain the outlier data of the sea ice, thereby highlighting the short-term fluctuations and seasonal changes of the sea ice; and the data is normalized and standardized.

3. The method for intelligent downscaling of polar sea ice density based on deep self-attention residual U-Net according to claim 1 is characterized by: In step 2, a data set for network training is constructed and divided into a training set and a test set; after the data is preprocessed, a data set for network training and verification is constructed; most of the data is divided into a training set according to a specific ratio for iterative training of the network; The smaller part is used as a test set to evaluate the fitting effect of the network model on non-training set data in real time and adjust the network parameters accordingly.

4. The method for intelligent downscaling of polar sea ice density based on deep self-attention residual U-Net according to claim 1 is characterized by: In step 3, the DeepSARU-Net network architecture is built. When setting the optimizer parameters, the main network of DeepSARU-Net is based on the encoding-decoding architecture of U-Net, which includes skip connections; during the downsampling operation, the strided convolution operation is used to replace the original maximum pooling operation of U-Net; at the same time, the square convolution kernel is decomposed into two mutually perpendicular rectangular convolution kernels to increase the depth of the network.

5. The method for intelligent downscaling of polar sea ice density based on deep self-attention residual U-Net according to claim 4 is characterized by: The DeepSARU-Net network architecture introduces an attention mechanism; The DeepSARU-Net network architecture is designed with multiple resolution stages. In the first jump connection with the final resolution stage, three key vectors are first introduced when constructing the mathematical framework of the attention mechanism: query tensor (Query, Q), key tensor (Key, K), and value tensor (Value, V). Among them, the query tensor Q represents the information currently being paid attention to and is the input of the attention mechanism; the key tensor K is used to measure the degree of correlation between the query tensor Q and the rest of the information; and the value tensor V carries the actual information associated with each query tensor Q. The mathematical expression of the attention mechanism is: In the above formula, d k Represents the dimension of K, and each row of the attention score matrix corresponds to the attention score calculated between the i-th row of the query vector Q and each column of the key vector K.

6. The method for intelligent downscaling of polar sea ice density based on deep self-attention residual U-Net according to claim 5 is characterized by: The attention scores are processed with a softmax operation to map the score values in each row to the interval [0, 1] and ensure that the sum of all elements in each row is strictly equal to 1.

7. The method for intelligent downscaling of polar sea ice density based on deep self-attention residual U-Net according to claim 6, characterized in that: In the convolutional neural network part, residual connections are introduced to optimize the gradient propagation path.

8. The method for intelligent downscaling of polar sea ice density based on deep self-attention residual U-Net according to claim 1 is characterized by: In step 4, the root mean square error (RMSE), correlation coefficient (r), peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are selected as evaluation criteria to evaluate the effect of DeepSARU-Net network downscaling, and the DeepSARU-Net network architecture parameters are reversely optimized based on the downscaling effect.

Images