【作者】 周琳；

【导师】 王树根； 邵振峰；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2022， 硕士

【摘要】 城市热岛已成为当今影响人类的主要气候问题之一,局部气候区分类体系为城市热岛研究提供了一个不同于传统城乡二分法的研究框架,将城市地表依据覆盖物性质、材料、形态特征等划分为17个类别,同一类别的气候区由于其地表覆盖、人类活动相似,因而具有相同或相似的气候响应。局部气候区可以为城市热岛研究提供更细粒度的地表信息,为城市热环境研究提供更丰富的数据支撑。局部气候区分类结果的准确性直接影响后续研究的结果,因此局部气候区分类也逐渐成为近年来的研究热点。基于遥感影像的分类方法是目前局部气候区分类的主流方法,这类方法面临着分类精度低、类别易混淆、样本标注复杂、难以跨域分类等问题。首先,局部气候区一些类别的特征非常相似,仅依靠光谱特征难以分辨,造成了分类结果的混淆,极大地影响分类精度。其次,由于不同城市所处气候背景、文化习惯不同,造成每个城市都有其独特的建筑形态特征,因此训练好的局部气候区分类模型很难直接用于其他城市的局部气候区分类。对于大范围的局部气候区分类,为每个城市都进行样本标注又是一项很难完成的任务,因此局部气候区跨域分类也是目前需要解决的一个问题。本论文重点面向局部气候区分类精度低、跨域分类难等问题,提出了针对性的方法,主要研究工作包括:(1)构建多源遥感影像城市局部气候区分类数据集。由于多光谱影像可以提供的信息有限,本论文使用北京、天津、武汉三个城市的Sentinel-1 SAR影像与Sentinel-2多光谱影像,共同构建了一个多源遥感影像城市局部气候区分类数据集,并命名为Sen12LCZ数据集,该数据集影像共包括19个通道,分别为SAR影像的9个特征通道与多光谱影像的10个通道。(2)提出基于深度学习的局部气候区分类方法。为了提高分类精度,提出了一种基于双注意力卷积神经网络与多源遥感影像的城市局部气候区分类方法,一方面使用具有强大特征提取能力的卷积神经网络挖掘数据的深层特征,在网络中加入通道注意力与位置注意力机制,抑制不重要特征的干扰,另一方面使用Sen12LCZ数据集,将多光谱影像与SAR影像结合,结合光谱特征、纹理特征、散射特征来提高局部气候区分类的精度。实验结果表明,多光谱影像与SAR影像的结合将分类的整体精度由0.9181提高至0.9425;与其他方法对比,本论文提出的局部气候区分类方法取得了更高的分类精度,该方法已发表于国际二区SCI期刊GSIS上。(3)提出基于深度领域自适应的局部气候区跨域分类方法。针对跨域分类问题,提出了一种基于全局与类别联合对齐的深度领域自适应算法,基于最优传输理论进行全局对齐,基于最大均值差异将源域城市数据与目标域城市数据在类别子空间进行对齐,实现局部气候区跨域分类。实验结果表明本论文提出的算法可以提高局部气候区跨域分类的精度,改善局部气候区跨域分类的效果。更多还原

【Abstract】 Urban heat island has become one of the major climate problems affecting human being today.Compared with the traditional urban-rural dichotomy,local climate zone provides a different research framework for the study of urban heat island.The urban surface is divided into 17 categories according to the material and morphological characteristics of the cover in the local climate zone scheme.Local climate zones of the same category have similar climate response because of similar surface coverage and human activities in these zones.Local climate zone can provide more fine-grained surface information for urban heat island research and richer data support for urban thermal environment research.The accuracy of local climate zone classification results directly affects the results of subsequent studies,thus local climate zone classification has gradually become a research hotspot in recent years.The classification methods based on remote sensing imagery are the mainstream methods of local climate zone classification at present.The research of local climate zone classification includes some key problems,such as low classification accuracy,easy confusion of categories,complicated sample annotation,and difficulty in cross-domain classification.Firstly,there are seventeen categories in the local climate zone system,and some categories are easy to be confused,such as compact high-rise buildings and compact middle-rise buildings,low vegetation and shrubs.It is difficult to accurately distinguish these categories only by using the spectral characteristics of multispectral images.Secondly,different cities are located in different climatic backgrounds,leading to unique building morphological characteristics in each city.Therefore,it is difficult to directly use the trained local climate zone classification model to classify local climate zones in other cities.For large-scale local climate zone classification,it is a difficult task to annotate samples for each city,so cross-domain local climate zone classification is also a problem that needs to be solved.This paper focuses on the problems of low accuracy of local climate zone classification and cross-domain classification,and proposes a targeted approach,the main research work includes:(1)An urban local climate zone classification dataset based on multi-source remote sensing images is constructed.Due to the limited information that provided by multispectral images,this paper uses Sentinel-1 SAR images and Sentinel-2multispectral images of Beijing,Tianjin and Wuhan,to jointly construct a multisource urban local climate zone dataset,which includes a total of 19 channels,9 channels of SAR images and 10 channels of multispectral images,respectively.(2)A deep learning-based method for local climate zone classification is proposed.In order to improve the local climate zone classification accuracy,a classification method based on dual attention convolution neural network and multi-source remote sensing images is proposed.On the one hand,convolution neural network with strong feature extraction ability is used to extract the deep features of data,and attention mechanism is added to the network to suppress the interference of unimportant features,On the other hand,multispectral images are combined with SAR images,and spectral features,texture features and scattering features are combined to improve the classification accuracy of local climate zone.The experimental results show that the combination of multispectral images and SAR images improves the overall classification accuracy from 0.9181 to 0.9425;compared with other methods,the proposed method achieves higher classification accuracy,and the initial method has been published on an SCI indexed journal GSIS.(3)A deep domain adaptive cross-domain classification method is proposed.Aiming at the problem of domain shift,a deep domain adaptive algorithm based on global and subspace joint alignment is proposed,which aligns the data of source domain with data of target domain in the global space and category subspace to realize the cross-domain classification of local climate zone.The results show that the algorithm proposed in this paper can improve the accuracy of cross domain classification of local climate zone,and can be used for large-scale local climate zone classification.更多还原