基于GF-4卫星影像时序光谱特征的居民地信息提取研究

发布时间：2019-03-24 15:37

【摘要】：随着“一带一路”与新型城镇化战略决策的推动与实施,中国在未来几十年中发展的空间格局将发生巨大改变。自1978年改革开放以来,我国的经济迅猛增长,在短时间内跃居全球前列,社会和谐发展,国民的生活质量有了明显的提升,居民地的扩张速度也愈来愈快。对居民地进行快速准确地识别和提取在推进国家战略决策,实现数字城市,辅助城市规划等多个领域具有重大的现实意义。高分四号(GF-4)卫星作为我国实施高分辨率对地观测系统的重要组成部分,能够及时有效识别地面变化,有效支撑地震、洪涝、干旱、台风等自然灾害救助、气候变化研究、林业及水资源环境调查等重大行业应用。本论文利用GF-4卫星影像,结合其高时谱这一特性,提取并分析居民地与其他地类的光谱特征差异,并结合时序光谱使用不同方法对居民地信息进行识别和提取。论文从以下几个部分展开:首先,介绍了本论文的研究背景和意义,再介绍与论文主题息息相关的遥感信息提取技术和居民地识别提取技术的研究现状与进展,并提出研究内容及技术路线。接着,介绍了GF-4卫星影像,并对影像进行预处理以消除来自各方面的误差。然后,通过对GF-4影像的典型地物光谱指数特征的分析,提出基于光谱特征决策树的居民地信息提取方法,进而对时序光谱指数特征进行分析,提出基于时序光谱特征决策树的居民地信息提取方法,在此基础上将时序光谱指数特征和深度学习技术同时引入居民地信息识别提取中,提出基于时序光谱特征全卷积神经网络的居民地信息提取方法。最终对三种方法的实验结果进行对比分析,得出结论。通过上述研究的开展,本论文可得到以下主要结论:(1)太阳高度角的变化不仅仅影响了地物光谱的大小,甚至对地物光谱的变化率大小和变化率变化的快慢也有一定的影响,且不同地物类型的光谱特征随太阳高度角的变化特征也有所不同。(2)当使用决策树方法时,结合时序光谱特征对居民地信息进行提取相较于仅结合光谱特征的提取来说,将提取精度由89.85%提升到93.38%。(3)利用全卷积神经网络可提升基于时序光谱特征居民地信息提取的提取精度,提取精度由93.38%提升到95.15%。此外,本论文有以下创新点:(1)将时序光谱特征与太阳高度角的关系引入到决策树模型中,相比较仅利用光谱特征的居民地提取方法而言,精度有所提高。(2)将时序光谱特征与深度学习中全卷积神经网络方法相结合,较不考虑时序光谱特征或不采用深度学习的其他提取方法来说,更加提升了分类提取精度。研究基于GF-4卫星遥感影像时序光谱的居民地识别提取方法,为减灾、防灾、推进城镇化进程、城市精细化管理和国土资源管理等工作快速提供动态更新数据,并为我国国产高分系列卫星数据遥感产品的应用提供技术与方法支撑和示范指导作用。
[Abstract]:With the promotion and implementation of "The Belt and Road Initiative" and the strategic decision of new urbanization, the spatial pattern of China's development in the coming decades will be greatly changed. Since the reform and opening up in 1978, the economy of our country has been growing rapidly, it has leaped to the forefront of the world in a short period of time, the harmonious development of the society, the quality of life of the people has been obviously improved, and the speed of the expansion of the residential land has also become more and more rapid. Rapid and accurate identification and extraction of residential land is of great practical significance in promoting national strategic decision-making, realizing digital city, assisting urban planning and so on. As an important part of China's high-resolution Earth observation system, the GF-4 satellite can effectively identify ground changes in a timely manner and effectively support natural disasters such as earthquakes, floods, droughts, typhoons, and so on. Climate change research, forestry and water resources environmental survey and other major industry applications. In this paper, we use GF-4 satellite image and its high-time spectrum to extract and analyze the difference of spectral features between residential land and other land classes, and use different methods to identify and extract resident land information combined with temporal spectrum. The thesis starts from the following parts: firstly, this paper introduces the research background and significance of this paper, and then introduces the research status and progress of remote sensing information extraction technology and residential identification extraction technology, which are closely related to the subject of the thesis. And put forward the research content and technical route. Then, the GF-4 satellite image is introduced, and the image is pre-processed to eliminate the errors from various aspects. Then, by analyzing the spectral index characteristics of typical ground objects in GF-4 images, a method of extracting resident land information based on spectral feature decision tree is proposed, and then the temporal spectral index features are analyzed. A method of extracting resident land information based on temporal spectral feature decision tree is proposed. On the basis of this method, temporal spectral index feature and depth learning technology are introduced into the identification and extraction of residential information at the same time. A method of extracting resident land information based on full convolution neural network based on temporal spectral features is proposed in this paper. Finally, the experimental results of the three methods are compared and analyzed, and a conclusion is drawn. The main conclusions of this paper are as follows: (1) the change of solar height angle not only affects the spectral size of the ground object, Even it has some influence on the spectral variation rate and the rate of change, and the spectral characteristics of different feature types vary with the solar height angle. (2) when the decision tree method is used, the spectral characteristics of the ground features vary with the solar height angle. (2) when the decision tree method is used, the spectral characteristics of the ground features vary with the solar height angle. Compared with the extraction of spectral features only, the time series spectral feature is used to extract the resident land information. The extraction accuracy is increased from 89.85% to 93.38%. (3) the extraction accuracy of resident land information based on temporal spectral features can be improved by using full convolution neural network, and the extraction precision is increased from 93.38% to 95.15%. In addition, the innovations of this thesis are as follows: (1) the relationship between temporal spectral features and solar height angle is introduced into the decision tree model, and compared with the resident extraction method which only makes use of spectral features, (2) combining the sequential spectral features with the full convolution neural network method in depth learning, the classification extraction accuracy is improved even more than other extraction methods which do not take into account the sequential spectral features or other extraction methods that do not use in-depth learning. (2) the sequential spectral features are combined with the full convolution neural network method in depth learning. Based on the temporal spectrum of GF-4 satellite remote sensing image, this paper studies the method of identification and extraction of land and land, and provides dynamic updating data for disaster reduction, disaster prevention, urbanization, urban fine management and land and resource management, and so on. It also provides technical and methodological support and demonstration guidance for the application of home-made high-grade series satellite data remote sensing products.
【学位授予单位】：中国科学院大学(中国科学院遥感与数字地球研究所)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P237

【相似文献】