Landsat卫星图像云和阴影去除算法研究

发布时间：2018-07-07 21:01

本文选题：云检测 + 阴影检测　；参考：《广西师范大学》2017年硕士论文

【摘要】：在信息时代的今天,遥感图像的采集和解译因计算机科学,电子信息,图像处理技术的进步而不断发展。大量的遥感资源图像具备高分辨率与广泛的区域覆盖特性,这些数据信息可使用在许多领域,如农业生产,环境监测,生态保护,基础建设投资,地形建模等。然而,受大气中不同形态的云和云阴影干扰,遥感卫星经常得到模糊的图像资源,严重影响了信息的时效性和完整性。另外在不同的背景下,亮白的云和较暗的阴影区域检测较困难,如果仅仅只使用人工方式检测云和云阴影,则极大影响信息分析的效率。为克服发展与效率之间的矛盾,自动化去云和云阴影成为研究遥感图像的一个很重要的科学领域。基于多幅不同时相遥感图像之间的信息补偿方式,提高了解译图像的清晰度和整体利用率。使用C++程序设计语言调用Opencv库函数,设计一套操作简单的图像处理软件,并成功处理广东和道县地区的遥感图像。论文的主要研究方法概括如下:文章采用传统的阈值、改进的分水岭、小波融合算法检测厚云区域。因厚云和薄云所具备物理属性不同,分别以不同的方式检测和处理。改进的分水岭算法先通过最大类间方差法标记亮白的区域,然后获得完备的厚云边界区域,避免了图像过度的分割。小波融合的方式充分利用蓝色通道图像厚云区域的完整性与近红外通道图像厚云与背景区域的差异性。多时相图像经3层小波分解,选取合适的小波分解系数,重构得到的图像厚云与背景区域对比度明显增强,利用传统阈值的方式对融合后的图像进行分割,能够得到灰度值不是很高的厚云区域。最后以上三种算法均通过区域膨胀的方式获得厚云边界上的薄云区域。在实际的应用中,为了排除灰度值相似于云的亮白背景对象和检测出灰度值相似背景的薄云区域,提出的掩膜算法,能够得到准确的云区域。遥感图像中的云阴影严重妨碍了信息提取和变化检测的精确度,城市建筑、山体、和云都能产生阴影,或者在光谱特性上与阴影相似的水体。本文为了排除水体并检测出云阴影,根据光谱间差异性检测出水体。在近红外波段下云的阴影比较暗,使用泛洪填充算法来获得可能的阴影区域。随后,通过几何形态学的方式构造云与云阴影之间的三维几何关系,确定云的阴影。为了减少云与阴影匹配的迭代次数,本文通过云的亮温值和光谱反射率确定云的高度,提高了阴影检测算法的有效性和准确性。为去除检测得到的云和云阴影区域,根据不同时相图像之间的差异,建立回归关系。使用矫正后的不同时相图像像素区域替换到被云和云阴影所污染的区域。从主观视觉的角度评估结果所示:上述云检测算法均能十分精确检测出云区域,其中掩膜算法效果最佳,最后得到无边界差异的去云效果图,并适合Landsat系列不同时相卫星图像资源。
[Abstract]:In the information age, the acquisition and interpretation of remote sensing images have been continuously developed due to the progress of computer science, electronic information and image processing technology. A large number of remote sensing images have high resolution and extensive regional coverage. These data information can be used in many fields, such as agricultural production, environmental monitoring, ecological protection, infrastructure investment, terrain modeling, and so on. However remote sensing satellites often get blurred image resources which seriously affect the timeliness and integrity of information due to the interference of cloud and cloud shadows in the atmosphere. In addition, it is more difficult to detect bright white clouds and dark shadow regions in different backgrounds. If only the manual detection of cloud and cloud shadow is used, the efficiency of information analysis will be greatly affected. In order to overcome the contradiction between development and efficiency, automatic cloud removal and cloud shadow has become an important scientific field in remote sensing image research. Based on the information compensation method of multi-time remote sensing images, the sharpness and overall utilization rate of the interpreted images are improved. Using C programming language to call Opencv library function, a set of simple image processing software is designed, and the remote sensing images in Guangdong and Daoxian areas are successfully processed. The main research methods are summarized as follows: traditional threshold, improved watershed and wavelet fusion algorithm are used to detect thick cloud region. Because thick cloud and thin cloud have different physical properties, they are detected and processed in different ways. The improved watershed algorithm uses the maximum inter-class variance method to mark the bright white region, and then obtains the complete thick cloud boundary area, which avoids the excessive segmentation of the image. The method of wavelet fusion makes full use of the difference between the integrity of thick cloud region of blue channel image and the difference between thick cloud and background region of near infrared channel image. The multi-temporal image is decomposed by three layers of wavelet, and the suitable wavelet decomposition coefficient is selected. The contrast between the thick cloud and the background area is obviously enhanced, and the fused image is segmented by the traditional threshold method. A thick cloud region with a low gray value can be obtained. Finally, the three algorithms are used to obtain the thin cloud region on the thick cloud boundary by the expansion of the region. In practical applications, in order to eliminate the bright white background objects with similar gray values and detect thin cloud regions with similar gray values, the proposed mask algorithm can obtain accurate cloud regions. Cloud shadows in remote sensing images seriously hinder the accuracy of information extraction and change detection. Urban buildings mountains and clouds can produce shadows or water bodies with spectral characteristics similar to shadows. In order to exclude water body and detect cloud shadow, the water body is detected according to spectral difference. In the near infrared band, the cloud shadow is dark, the flooding fill algorithm is used to obtain the possible shadow region. Then, the geometric relationship between cloud and cloud shadow is constructed by geometric morphology, and the cloud shadow is determined. In order to reduce the number of iterations of cloud and shadow matching, the cloud height is determined by cloud brightness temperature and spectral reflectivity, which improves the effectiveness and accuracy of shadow detection algorithm. In order to remove the detected cloud and cloud shadow regions, a regression relationship was established according to the differences between different phase images. The corrected pixel area of the different phase image is replaced by the area contaminated by cloud and cloud shadow. The evaluation results from subjective vision show that the above cloud detection algorithms can detect cloud region accurately, and the mask algorithm has the best effect. Finally, the de-cloud effect map without boundary difference is obtained. And suitable for Landsat series of different phase satellite image resources.
【学位授予单位】：广西师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【参考文献】