【作者】 周淑媛；

【导师】 冯学智； 肖鹏峰；

【作者基本信息】 南京大学 ， 地图学与地理信息系统， 2016， 硕士

【摘要】 积雪作为冰冻圈的重要组成部分之一,在全球气候变化、地表辐射传输与平衡扮演着重要的角色。同时积雪也是宝贵的淡水资源库,对区域水循环、气候变化、水文调查研究有着重要的作用。积雪识别研究为雪水当量估算、融雪径流模拟与预报、雪灾评估、区域水资源管理等提供了基础。合成孔径雷达(Synthetic Aperture Radar,SAR)技术使得大面积重复观测、全天时全天候获取区域尺度的积雪信息称为可能,在气候条件恶劣的环境下,该技术已成为一种重要的对地观测手段。极化SAR技术可获得不同极化方式下的地物散射信号,为积雪识别提供了更为丰富的信息。如何有效利用极化SAR图像优势,克服其受相干斑噪声和地形影响的缺点进行积雪信息识别是目前亟需解决的问题。本文依据国家自然科学基金项目“SAR与高分辨率光学遥感联合反演雪水当量”(项目编号41271353)对SAR图像积雪识别的研究需求,以新疆天山北坡玛纳斯河流域山前坡地为研究区,选择积雪期与融雪期两景极化RADARSAT-2数据,对极化SAR图像积雪识别理论与方法进行了探讨与研究,着重研究了积雪的微波特性和SAR图像表征,并结合图像空间关系和先验信息,克服相干斑噪声影响,建立积雪识别模型,为极化SAR图像积雪识别提供新思路。论文的主要研究内容与结论包括:(1)积雪的微波特性与SAR图像表征分析。积雪的微波特性包括散射特性和极化响应特征,其微波特性受积雪的物理特性和SAR系统的参数影响。不同积雪表面干湿状态下的散射信号不同,对于干雪而言,C波段SAR数据可穿透一定厚度的干雪,主导散射为雪-地界面的散射;对于湿雪而言,C波段SAR数据穿透能力减弱,主导散射为雪-空气界面的散射。通过积雪期和融雪期后向散射系数图和极化总功率图表征分析可知,积雪的后向散射系数和极化总功率均低于研究区内其他地物;同时融雪期的积雪后向散射系数比积雪期下降了约10dB,积雪在交叉极化方式下与研究区其余地物可分性较高;极化总功率综合了四种极化方式下的散射特性,可分性也较高。极化SAR数据测量得到的复散射矩阵包含了各极化通道下的散射信号及各通道之间的关系,其多视协方差矩阵的统计模型可由复Wishart分布、K-Wishart分布等进行表达。(2)基于马尔可夫随机场的极化SAR积雪识别模型构建。针对SAR图像受相干斑噪声影响的现象,利用MRF模型结合空间上下文信息,充分利用图像像元之间的关系,结合先验信息和极化SAR图像自身特征。探讨多视极化SAR数据满足的分布函数,验证了 K-Wishart分布可表达多视极化SAR数据的分布情况,有较高的拟合度。通过贝叶斯理论建立积雪识别模型,通过图像先验概率与概率密度函数求解积雪与非积雪标记场的最大后验概率,首先利用Wishart距离进行模型初始化,获取初始标记场,然后进行模型参数估计,利用ICM算法迭代求解最大后验概率,直至模型收敛,获取最终的积雪最优识别结果。(3)积雪识别结果分析。利用基于MRF的极化SAR积雪识别模型对研究区两景SAR图像进行积雪识别,通过地形与识别结果叠加图结合实测数据分析积雪识别结果,并计算准确率(Precision)、召回率(Recall)和F指数(F-score)定量评价识别结果。结果表明:经实测数据验证,通过结合空间上下文信息的K-Wishart分布MRF积雪识别模型,相对于其他分布模型和未利用MRF模型的积雪识别模型在识别精度、结果的完整性和对噪声的敏感性具有较大优势,识别结果孤立点较少,边缘较为平滑且准确,表明了该方法识别积雪的可靠性。论文基于极化SAR图像,在分析积雪期与融雪期积雪的微波特性和SAR图像表征基础上,针对SAR图像受相干斑噪声影响的缺点,利用MRF模型结合图像空间上下文信息和先验知识,探讨合适的SAR数据分布模型,建立积雪识别模型,为解决SAR图像受相干斑噪声影响的问题提供了新思路,具有一定的理论和应用创新性。更多还原

【Abstract】 Snow cover is an important component of the cryosphere.It plays an important role in global climate change,regulating surface radiation energy budget and balance.As freshwater resource,snow cover is very significant for regional water cycle,climate change,and hydrological investigation research.Recognition of snow provides a basis for snow water equivalence estimates,snowmelt runoff simulation,snow disaster evaluation,and water resources management.Synthetic Aperture Radar(SAR)technology,which has high spatial resolution,can obtain the large range snow cover information in all-weather and all-time,especially for the poor weather conditions and high-altitude cold regions.The polarimetric SAR(PolSAR)can get the scattering information of ground objects with different polarization modes,which provides more information for the snow cover recognition.How to overcome the influence of speckle noise and topography is an urgent problem.This paper focuses on the need of National Natural Science Foundation "Joint inversion of snow water equivalence based on SAR and high-resolution optical remote sensing"(Grant No.41271353).The full polarimetric RADARSAT-2 image in snow accumulation and snow melt period is utilized in this study.The study area is located in Manasi River basin of Tianshan Mountain.First,the microwave characteristics and the image representations of snow are discussed.Then,the snow recognition algorithm based on Markov Random Field(MRF)model is proposed,which uses the spatial-contextual information and prior knowledge.Finally,this method is employed to extract snow cover area from PolSAR images.The main contents and conlusions are as follows:(1)Analysis of the microwave properties and SAR image characterization of snow cover.The microwave characteristics are affected by the physical properties of snow and the parameters of the SAR system.For dry snow cover,the main backscattering is surface scattering from the snow-ground.In wet snow,the absorption loss is high and the dominate backscattering is surface scattering of air-snow.Research shows that the backscattering coefficients and the total scattered power both are lower than other surface features of the study area.The backscattering coefficient of snow accumulation period declined about 10 dB compared with snow melt period.According to the characterization analysis,the backscattering coefficients in cross polarization mode and the total scattered power have a good ability to distinguish the snow and snow-free.The polarimetric scattering matrix contains scattering signal and relationship of four polarization channels,which has more abundant information than single polarization data.The statistical model of multilook complex covariance matrix data can be expressed by complex Wishart distribution or K-Wishart distribution.(2)Construction of the snow cover recognition of polarimetric SAR image besed on Markov Random Field.In order to reduce the effect of spekle noise,a novel recognition model of snow cover is proposed using MRF.MRF can take advantage of spatial contextural information,so that it can reduce the effect of speakle noise.The algorithm uses the K-Wishart distribution for the PolSAR data statistics model,because it has a good fit on the real PolSAR data.In this method,the PolSAR image is initialized by Wishart distance algorithom at first.Then,the model parameter is estimated by iteration and the maximum a posteriori(MAP)is sloved by iterative conditional mode(ICM)algorithom.Until the model is convergenced,the optimal snow recognition results are acquired.(3)Analysis of the recognition result of snow cover.The proposed model based on MRF,is used to recognize snow cover from PolSAR images.Using the measured ground truth data verifies the results of snow cover recognition.What’s more,the quantitative indexes,Precision,Recall and F-score,are calculated to evaluate the recognition results.Compared with the other methods,the method based on MRP model combined with a K-Wishart distribution for the PolSAR data statistics model achieves higher accuracy and better results.This recognition model can obtain the homogeneous and reliable snow cover results even on the SAR data with a high degree of speckle.Aiming at the disadvantage of SAR image with speckle noise,this study proposed the model of snow cover recognition based on the MRF model,combining the statistical a priori knowledge with the the spatial contextual information,which can obtain the exact and complete snow cover results.It provides a new idea for snow cover recognition with SAR image,and has a certain theoretical and applied innovation.更多还原