【作者】 王龙；

【导师】 王萍；

【作者基本信息】 天津大学 ， 控制科学与控制工程， 2014， 硕士

【摘要】 强对流天气指的是发生突然、变化剧烈、破坏力极强，常伴有雷雨大风、冰雹、龙卷风、局部强降雨等强烈对流性灾害天气。在多普勒天气雷达反射率图中，此类天气状况的发生与强对流风暴单体的相关性显著。在进行强对流预报工作中，云团的多普勒雷达反射率图是十分重要的参考资料，当前预报的主要模式也是基于对风暴体运动变化的趋势外推来进行的。为了在外推时尽可能的保留强对流风暴单体多普勒雷达图像结构与形态信息，本文提出一种“分层-双向膨胀-合并-旋转”(RDBCE)外推算法。该算法以VS2010MFC和OpenCV为开发工具，综合运用计算机视觉、图像识别技术和气象科学知识，设计强对流风暴单体的自动识别和外推模型，主要工作可以分为以下几个方面：首先读取相邻时刻风暴单体图像，判断出其位置偏移和角度偏移，根据强对流风暴体的层次结构特点，将单体的彩色图像分解成多幅二值图像；对每幅图像，识别出生长区域和收缩区域；提出对生长区域正膨胀和收缩区域反膨胀的双向膨胀算法，解决图像外轮廓凹凸变化的外推；提出利用拟合椭圆长轴方向变化解决跟踪对象的旋转信息并实现旋转外推；各外推子图像合成后，恰当地延续了不同图像区域的变化趋势，使结构和形态信息得以保留。对于多个单体交叠的大块风暴云团，该算法也设计了单体分别外推再融合到一起的模型，克服了传统TREC算法中只考虑风暴体整体及局部位移的不足，为多核风暴云团结构及形态外推提出了新的思路。实验结果表明，本文外推图像与真实图像的几个重要特征无显著性差异。样本的6分钟外推图像与真实图像相似度达到70%，其中多核风暴云团样本的相似程度较业务中普遍应用的TREC算法平均提高了8个百分点。更多还原

【Abstract】 Convective weather refers to a sudden and extremely destructive weather, which isoften accompanied by thunderstorms, hail, tornadoes, strong local convective rainfalland other weather disasters. Doppler weather radar reflectivity figure shows that thereis an obvious correlation between severe convective storm cell the occurrence of suchweather conditions. In nowadays work, Doppler radar reflectivity image of cloud is avery important data to make forecast, and the main mode of the current convectiveweather forecast is also based on storm cell tracking and extrapolating.To retain the morphological characteristics of storm cell as much as possible anextrapolation algorithm called rotation, decomposition, bilateral dilation andcomposition extrapolation (RDBCE) is presented to obtain an extrapolated Dopplerradar reflectivity image. The algorithm was coded in VS2010MFC and OpenCV,integrated use of computer vision, image recognition technology and meteorologicalscientific knowledge, convective storm cell automatic identification and extrapolationmodel was designed as following:First, get2storm cells of consecutive time and obtain the position offset andangular offset. Based on the hierarchical structure of a storm cell, a color reflectivityimage is first decomposed into multiple binary images. By matching the currentimage with its predecessor, expanding and shrinking areas can be identified. Bilateraldilation algorithm is then implemented to predict changes in image contours.Meanwhile the rotation and orientation prediction of a storm cell is represented by anellipse stretching along its long axis. After the composition, rotation and translation ofsub-images, a final extrapolated Doppler radar reflectivity image which retains thevariation trends and morphological characteristics is generated.For storm cloud which is composed of multiple overlapping cells, we designed amerging extrapolation model. Unlike the traditional TREC algorithm using local areachanges to represent internal morphological changes, the algorithm introduced in thispaper purposes an innovative approach to forecast morphological changes of a stormcell. Test results show that the similarity rate between6minutes-extrapolated imageand the actual image is up to70%，among which storm clouds with multiple cores isimproved by8%on average compared to the TREC algorithm used in operationalsystems. Furthermore, no significant discrepancy has been identified betweenextrapolated image and actual image.更多还原