随着全球变暖，北极海冰的面积呈现下降趋势，日益成为科学家关注的焦点。虽然北极地区地处偏远，但是作为地球系统的一个重要组成部分，该地海冰变化对气候、环境、生物多样性等的影响巨大。自20世纪下半叶有卫星观测开始以来，海冰观测就逐渐受到广泛关注。特别是SAR卫星相继发射以来，海冰观测多为卫星设计与应用中重要的目的之一。 本文在前人研究基础上，以西北航道作为研究区域，利用Sentinel-1 A/B SAR数据，基于灰度共生矩阵和LibSVM分类方法将2015年11月-2018年2月三个冬季97景西北航道影像数据内的海冰划分为一年冰和多年冰。 数据处理方面，Sentinel-1 A/B是欧空局最新一代SAR卫星，其数据获取方式采用了德国宇航中心提出的TOPSAR技术，它能够在保证分辨率较高的同时扩大地面的覆盖范围，但是与ScanSAR模式相同，获得的影像中的热噪声仍较为严重。因而，在对数据进行预处理过程中，需对影像的热噪声进行去除。本文采用SNAP软件中提供的热噪声移除模块对于热噪声进行了去除，一定程度上抑制了热噪声对于分类结果的影响。其次，在Sentinel-1 A/B数据在从Level-0级转化为Level-1级的过程中，影像的方位向边缘出现边缘噪声，本文基于15景影像中间行后向散射值的统计结果，选择-40dB作为阈值对边界噪声予以去除。此外，在预处理过程中解决的另一个关键问题是SAR数据由侧视观测而产生的角度效应问题，基于15景数据一年冰与多年冰后向散射值与入射角之间的相关性，对于该效应进行了校正。除此之外，在预处理过程中，还进行了精密轨道校正、定标转换、投影转换以及陆地掩膜操作。 方法方面，本研究划分为两部分，即基于灰度共生矩阵(Grey Level Co-occurrence Matrix, GLCM)提取影像特征和基于LibSVM的海冰分类。研究中将15景数据作为训练样本，利用GLCM获取SAR数据的10种纹理特征。在提取纹理特征时，GLCM有四个重要输入参数，分别是灰度级、窗口大小、位移距离以及原始窗口与目标窗口的位移方向。通过设置不同的参数组合获取不同的纹理特征，然后目视判读获得最优的参数组合为灰度级32，窗口大小为7×7个像元，步长为3个像元，选择0度方向作为窗口的位移方向。由于10种纹理特征数量较多，在利用分类器分类前，本研究结合纹理特征的散点图进行筛选，最终选择平均值、方差的对数、偏斜度绝对值的对数这3个纹理特征作为分类参数。在LibSVM分类器中，选择采用RBF核函数，通过交叉验证确定该分类方法中最重要的两个参数是惩罚系数C与松弛变量γ，分别取32768与2。随后将训练好的模型应用于97景SAR数据，得到一年冰和多年冰的分类结果。 精度验证方面，本文采用加拿大冰服务中心(Canadian Ice Service，CIS)的矢量冰图作为验证数据，该数据以划分多边形的方式存储每个区域内的多种海冰类型。由于其分类较多，而本文的目标仅将海冰分为一年冰与多年冰两种类型，因而对海冰类型进行了重归类。由于每一多边形混合有一年冰与多年冰两类而无法对应于相应的地理位置，因而在重归类的栅格化过程中需要解决的另一个问题即每一多边形内海冰类型的确定问题。本研究选择50%这一比例作为阈值进行栅格化操作，即将占比超过50%的海冰类型作为该多边形的海冰类型而忽略占比未过半的海冰类型，当两者的占比相同时，选择SA标签所记录的海冰类型作为该区域内海冰类型，由于CIS采用依海冰厚度从厚到薄的方式依次对海冰进行记录，因而SA记录的即为多年冰。精度评价结果显示，97景影像分类精度的加权平均值为73.57%。其中，除2015年11月与2016年12月两个月份的分类精度低于70%，其他月份的分类精度均在70%以上。通过将本文的分类结果与CIS冰图分类结果进行对比分析，研究发现，CIS提供的解译冰图因为时间间隔较长且因冰图同一多边形区域内不是单一类型的海冰，影响了分类结果的精度。因而，精确地面验证数据的缺乏是海冰分类面临的一个重要问题。同时，Sentinel-1A/B SAR数据中存在多种噪声与角度效应，未来精确的海冰分类需要对其数据质量问题进行很好地解决。

With global warming proceeding, sea ice extent in the Arctic has been increasingly decreasing and Arctic sea ice has become the center of attention for scientists all over the world. As an important component of the earth system, the Arctic is very significant though it is in remote area. The influence of the reduction of sea ice to climate, environment and biodiversity cannot be neglected. Since the beginning of satellite observations in the last century, sea ice has been highly concerned. Especially since the successive launch of SAR satellites, sea ice observations have been one of the important purposes in satellite design and application. This paper uses Grey-Level Co-occurrence Matrix(GLCM) to extract textural features for 97 Sentinel-1A/B SAR imageries during winter from November 2015 to February 2018 in Northwest Passage and LibSVM method to classify sea ice to First-year Ice(FYI) and Multi-year Ice(MYI). In terms of data processing, Sentinel-1 A/B SAR data is the latest generation of SAR satellite data for ESA. The data acquisition method adopts the TOPSAR technology proposed by the German Aerospace Center(DLR) ,which can improve the coverage of the ground with ensuring the resolution. However, serious thermal noise in imagery is similar to the ScanSAR mode. Therefore, noise removal of imagery is required in data pre-processing process. Thermal noise removal module provided in the SNAP software is used to remove noise in this paper, which weakens the influence of thermal noise on classification result. Besides, edge noise occurs at the azimuth direction boundaries during Level-0 to Level-1 process. -40 dB is selected as the threshold of edge noise removal based on the middle line of 15 images. Angular effect of SAR data from side-view observation is another important issue need to be addressed during the pre-processing process. Angular effect correction is performed based on the correlation between FYI and MYI backscatter values and the incident angle of 15 images. Precise-orbit correction, scaling conversion, projection conversion and land mask operations were also performed in the pre-processing process. In terms of methodology, there are two parts, one is textural features based on GLCM and another is classification using LibSVM classifier. 15 imageries are used as training samples and 10 textural features of SAR data are obtained using GLCM. Four parameters are significant for GLCM acquisition for every images, include grey level, displacement distance, window size, and displacement direction. Different parameters combination produce diverse textural features. The optimal parameters are determined by the visual interpretation, grey level is 32, displacement distance is 3, window size is 7 pixels. Displacement direction using 0 degree direction. Due to 10 textural features are redundant, mean value, logarithm of variance value and logarithm of absolute value of skewness is filtrated and selected as parameters to use in classifier referred to different textural features of scatter figures before classification. RBF kernel function is used in LibSVM classification method and penalty coefficient(C) and smoothness parameter(γ) are determined using cross-validation, respectively 32768 and 2. Then model trained best was applied to 97 SAR imageries and get FYI and MYI classification result. In terms of precision validation, Canadian Ice Service (CIS) vector ice charts is the validation data. Every polygon in CIS charts is a mixture of multiple sea ice types. Due to its too many types and the objective of our study is to classify sea ice into two types, FYI and MYI, so reclassification is necessary. 50% proportion of the polygon is the threshold to re-classified in rasterization procedure, that means the sea ice type exceed half is set in the polygon. Sea ice types in SA label will be used in those polygons which FYI and MYI are both 50% percentage. Due to sea ice types are stored according to thickness in CIS egg code figure, and SA tag definitely record MYI in re-classification result. Weighted average value of all 97 imageries classification result is 73.57% based on CIS ice chart. The classification accuracy of all months is above 70% except November 2015 and December 2016. We found that temporal resolution of CIS ice charts are a bit long and not one single ice types in every polygon, that influence classification. Besides, shortage of precise ground validation data is a significant problem for sea ice classification. Meanwhile, it need to be solved well that multiple noise and angle effect in Sentinel-1 A/B SAR high precision classification work in the future.