海洋环流与波动重点实验室知识库

受全球气候变化影响，北极海冰在范围、厚度和年龄上都有明显的减少趋势，大量永久性多年冰（Multi-year ice, MYI）被季节性一年冰（First-year ice, FYI）取代。MYI不但与北极夏季最小海冰范围密切相关，同时也是威胁北极船只安全的重要因素。波弗特海是北极MYI减少最严重的区域之一，也是北极MYI的主要输出区域之一，还是北极西北航道的重要组成部分。因此研究波弗特海区域的高分辨率海冰分类模型，并进一步研究其MYI变化特征对MYI监测和北极航线规划具有重要意义。

合成孔径雷达（Synthetic Aperture Radar, SAR）有全天时、全天候监测能力，是常用的高分辨率海冰监测传感器。受益于人工智能的快速发展，基于SAR的海冰分类从传统阈值法发展到机器学习和深度学习方法，分类精度和效率有了显著提升。然而仍有几个问题等待解决:（1）现有深度学习方法多采用卷积神经网络（Convolutional Neural Network, CNN），其不是一个像素级的分类方法，且分类准确率有待提升；（2）现有深度学习模型侧重于采用SAR图像的极化信息作为分类依据，对SAR图像的纹理特征存在忽视，限制了分类精度的提高；（3）波弗特高压变化会影响波弗特海域MYI的漂移，近年来，波弗特高压异常事件时有发生，在此背景下，波弗特海MYI的变化呈现的特征变化，有待进一步揭示。为解决上述问题，本文以波弗特海为研究区域，Sentinel-1 SAR影像为数据源，构建了基于深度学习的海冰智能化分类模型。进一步利用分类模型获取2018-2022年波弗特海的MYI变化特征，分析了其影响因素。具体为以下三方面内容：

1. 搭建了基于U-Net的像素级SAR图像海冰分类模型。本文首先收集了24幅Sentinel-1在EW模式下获取的双极化SAR影像，利用目视解译、参考MYI产品和时序数据验证方法标注真值。利用16幅图像构建训练集，剩余8幅构建测试集。引入深度学习U-Net模型，搭建了像素级的海冰分类模型，分类SAR影像中的MYI、FYI和开放水域（Open Water, OW）。U-Net模型在测试集取得了较高的精度，在准确率（Acc）、平均重叠率（mIoU）和Kappa系数分别取得了90.2%、0.796和0.817。本文开展了模型对比和敏感性测试实验。实验结果表明，U-Net模型精度远高于现有的海冰分类模型，如：支持向量机（Support Vector Machine, SVM）、随机森林（Random Forest, RF）和CNN。深度学习模型在HV去噪后的分类结果与去噪前相差较小，相较于传统的机器学习模型，深度学习模型具有一定的抗噪声能力。U-Net模型中，单一斜率校正HH图像对分类FYI和MYI有利。单支输入模式下，在双极化信息基础上加入单一灰度共生矩阵（Gray-level co-occurrence matrix, GLCM）纹理特征，能提升模型对OW的分类能力，但对FYI和MYI分类精度的提升不明显。以上研究表明深度学习模型U-Net在海冰分类较其他模型有明显优势。

2. 搭建了融合SAR图像极化信息和纹理特征的DBU-Net海冰分类模型。为进一步提高U-Net的海冰分类精度，本文利用双分支编码器和6种GLCM纹理特征改进U-Net模型，构建了双分支U-Net（DBU-Net）海冰分类模型。双分支编码器能更好地融合SAR图像的极化信息和纹理特征，使DBU-Net的海冰分类精度在Acc/MIoU/Kappa达到92.0%/0.846/0.851，比原始U-Net模型提高1.7%/5.0%/3.4%。本文对DBU-Net开展了模型对比和敏感性测试实验，并利用不来梅大学发布的MYI产品（Bremen MYI）验证DBU-Net分类结果。实验结果表明，在模型输入为极化信息和GLCM纹理特征时，DBU-Net模型精度远超SVM、RF和CNN。双分支结构和GLCM纹理特征的敏感性实验结果表明，两者对模型精度提升均有帮助，明显提高了模型对OW和FYI的分类能力。在与现有MYI产品的比较中，DBU-Net分类结果和Bremen MYI在波弗特海和弗拉姆海峡均表现出较小差异。以上研究验证了DBU-Net海冰分类的鲁棒性，为生成波弗特海海冰分类结果提供了技术支持。

3. 研究了2018-2022年波弗特海冬季多年冰变化特征。本文收集了415幅获取自波弗特海域2018年10月至2022年3月的Sentinel-1 SAR图像，利用训练完善的DBU-Net生成了60个基于SAR的海冰数据（SAR-based Sea Ice Data, SSID），获取波弗特海FYI和MYI的变化特征。并进一步利用海冰漂移数据和SLP数据分析两个典型的波弗特海MYI变化特征和影响因素。得到结论如下：（1）冬季波弗特海MYI的总体变化特征可分为三种。MYI总体输入，如：2019/20和2020/21冬季；MYI总体输出，如：2018/19冬季；和MYI总量几乎不变，如：2021/22冬季;（2）导致2018/19年冬季的MYI输出的因素是波弗特高压引起的海冰漂移异常，从通常的东北向西南漂移变为东南向西北漂移;（3）导致2020/21年冬季的MYI输入的因素是波弗特高压引起的强烈反气旋波弗特环流，带动了大量北向南的MYI输入；同时初始MYI较少，推迟了MYI漂移至西边界的时间，从而减少了西边界输出。

综上所述，本文设计并搭建了DBU-Net高分辨率海冰分类模型，精确分类SAR影像中的MYI、FYI和OW。进一步利用模型和SAR图像生成了60幅2018-2022年冬季波弗特海SSIDs，分析了两个典型的MYI变化特征和影响因素。研究结果为进一步开发整个北极的海冰分类模型和生成波弗特海MYI分布特征提供了技术支持。

As a result of global climate change, the Arctic sea ice is decreasing significantly in extent, thickness, and age with large amounts of permanent Multi-Year Ice (MYI) being replaced by seasonal First-Year Ice (FYI). MYI is not only closely related to the Arctic summer sea ice minimum extent, but also an important factor affecting the navigation of the Arctic. The Beaufort Sea is one of the areas with the greatest reduction in MYI and one of the main export areas of the Arctic MYI, as well as being an important part of the Northwest Passage in the Arctic. Therefore, the development of a high-resolution sea ice classification model for the Beaufort Sea and further study of its MYI change characteristics are of great significance for MYI monitoring and Arctic navigation.

Synthetic Aperture Radar (SAR) is a commonly used high-resolution sea ice monitoring active sensor with weather-independent monitoring capability. Benefiting from the rapid development of artificial intelligence, SAR-based sea ice classification has evolved from the traditional threshold method to machine learning and deep learning methods. The classification accuracy and efficiency have been significantly improved. However, there are still several problems waiting to be solved: (1) Most existing deep learning methods use Convolutional Neural Networks (CNNs), which are not pixel-level classification methods and there is potential to improve the classification accuracy. (2) Existing deep learning models only use SAR polarization data as the reference. Texture features that are helpful in traditional machine learning classification models are not addressed. (3) The Beaufort High variation affects the sea ice motion in the Beaufort Sea, which further affects the import and export of MYI. Therefore, the characteristics of MYI variability in the Beaufort Sea under the influence of variable Beaufort High need further study. To address the above problems, this thesis develops and improves a sea ice classification model based on U-Net with the Beaufort Sea as the study area and Sentinel-1 SAR images as the data source. Furthermore, the model is used to obtain the characteristics of the MYI variation in the Beaufort Sea from 2018-2022, and the factors are analyzed. This research mainly carries on the following three tasks.

1. A pixel-level U-Net-based sea ice classification model with SAR images. Firstly, we collected 24 Sentinel-1 dual-polarized SAR images acquired in EW mode. Visual interpretation, MYI products, and time-series SAR data are used to label the ground truth. Sixteen images are used to construct the training set and the remaining eight to construct the testing set. A deep learning U-Net model is introduced to build a pixel-level sea ice classification model to classify MYI, FYI, and Open Water (OW) in SAR images. The U-Net model achieves high accuracy in the testing set, with 90.2%, 0.796, and 0.817 in accuracy (Acc), mean intersection over union (mIoU), and Kappa coefficient respectively. Model comparison and sensitivity experiments are carried out. The experimental results show that the accuracy of the U-Net model is much higher than that of existing sea ice classification models, such as Support Vector Machine (SVM), Random Forest (RF), and CNN. The deep learning model has noise immunity compared with the traditional machine learning model. A single slope-corrected HH image in the U-Net model is beneficial for classifying FYI and MYI. In the single-branch input mode, the addition of single Gray-Level Co-occurrence Matrix (GLCM) texture feature on the basis of the dual-polarization information can improves the model's ability to classify OW, but has no improvement on FYI and MYI. These studies show that the deep learning model U-Net has a more significant advantage than other models for sea ice classification.

2. A DBU-Net sea ice classification model fusing polarization information and texture features. To further improve the sea ice classification accuracy of U-Net and reduce the misclassification in the testing set. The U-Net is improved to the Dual-Branch U-Net (DBU-Net) using a dual-branch encoder and 6 types of GLCM texture features. The dual-branch encoder can better fuse the polarization information and texture features, improving sea ice classification evaluation metrics of DBU-Net to 92.0%/0.846/0.851 in Acc/mIoU/Kappa, which is 1.8%/5.0%/3.4% better than the original U-Net model. Model comparison and sensitivity experiments are carried out on the DBU-Net. Furthermore, the DBU-Net classification results are validated using the MYI product published by the University of Bremen (Bremen MYI). The experimental results show that the DBU-Net model is more accurate than SVM, RF, and CNN with polarization information and GLCM texture features as the model inputs. The sensitivity experiment on the dual-branch structure and GLCM texture features shows that both of them contribute to the model accuracy improvement, significantly improving the model's ability to classify OW and FYI. In comparison with existing MYI products, DBU-Net classification results and Bremen MYI show small differences in both the Beaufort Sea and Fram Strait. The above study validates the robustness of DBU-Net in sea ice classification and provides technical support for the generation of regional sea ice classification results.

3. Characterization of winter MYI variability in the Beaufort Sea 2018-2022. 415 Sentinel-1 SAR images are collected which are acquired from the Beaufort Sea from October 2018 to March 2022. The well-trained DBU-Net is used to generate 60 SAR-based Sea Ice Data (SSID) achieving the trend of FYI and MYI in the Beaufort Sea. Furthermore, the sea ice motion data and Sea Level Pressure (SLP) data are used to analyze the factors contributing to the MYI variation characteristics in the Beaufort Sea. The following conclusions are obtained: (1) The overall trend of MYI during winter can be categorized into three states: import state, the winters in 2019/20 and 2020/21; export state, the winter in 2018/19; and minor motion state, the winter in 2021/22. (2) Factors contributing to MYI export in winter 2018/19 were anomalous sea ice drift caused by Beaufort High, which changed from the usual northeast to southwest sea ice motion to the southeast to northwest motion; (3) Factors contributing to MYI import in winter 2020/21 were the strong anticyclonic Beaufort Grye caused by Beaufort High, which drove a large amount of north-to-south MYI import; additionally, the less initial MYI delayed the MYI drift to the western boundary, thus reducing the export on the western boundary.

In summary, we design and develop a DBU-Net high-resolution sea ice classification model to accurately classify MYI, FYI, and OW in SAR images. Additionally, the model and SAR images are used to generate 60 SSIDs of the Beaufort Sea for winter 2018-2022, and the trends and influencing factors of two typical MYI states are analyzed. The results of the study provide technical support for the further development of a sea ice classification model for the whole Arctic and the generation of MYI distribution characteristics for the Beaufort Sea.