出版日期: 2016-11-25点击次数：下载次数：DOI: 10.11834/jrs.201652342016 | Volumn20 | Number 6 上一篇|

1. 南京信息工程大学 应用气象学院，南京 210044
2. 中国科学院遥感与数字地球研究所遥感科学国家重点实验室，北京 100101
 收稿日期: 2016-01-22; 修改日期: 2016-05-18; 优先数字出版日期: 2016-11-25 基金项目: 中国科学院重点部署项目(编号: KZZD-EW-TZ-18)；国家自然科学基金(编号: 41371360) 第一作者简介: 李彬(1989—)，男，硕士研究生，研究方向为地表辐射平衡。E-mail： 983229508@qq.com 通讯作者简介: 辛晓洲(1976—)，男，副研究员，研究方向为地表辐射与能量平衡遥感估算理论与方法。E-mail： xin_xzh@163.com 中图分类号: TP79 文献标识码: A

# 关键词

Cloud edge heights matching between thermal infrared and visible data
LI Bin1,2 , XIN Xiaozhou2 , ZHANG Hailong2 , HU Jichao1
1.School of Applied Meteorology, Nanjing University of Information Science & Technology, Nanjing 210044, China
2.State Key Laboratory of Remote Sensing Science, Institute of Remote Sensing and Digital Earth, Chinese Academy of Sciences, Beijing 100101, China

# Abstract

Differences in cloud height are obvious in high-resolution data, especially because cloud edge heights have become an important factor in cloud shadow identification and estimation of surface solar radiation. However, the resolution of cloud heights calculated by thermal infrared data is low and lacks detailed characteristics. Cloud edges of visible and thermal infrared bands differ considerably both in shape and geometric features. The edge of high-resolution image has rich characteristics, whereas that of thermal infrared cloud height data is single and fuzzy in geometric characteristics, so they cannot match exactly. Although some feature points of clouds can be obtained by some feature point matching methods such as Scale-Invariant Feature Transform (SIFT) and Harris, the difference between two data on geometric features made available by feature points was less. This result cannot satisfy the need to match the information of thermal infrared cloud heights and high-resolution cloud edge data, and obtain a result with abundant diversity. To solve this problem, an algorithm was presented in this study. First, SIFT algorithm was utilized in this method to extract feature points for further image registration and correction. Then, cloud edge heights were calculated by thermal infrared data and re-sampled to a high resolution. Next, Euclidean distance transform was performed for each cloud edge pixel of high-resolution data to all thermal infrared cloud edge pixels, which could obtain spatial relationships between the two types of data. As the two types of data differed considerably in edge characteristics, directly determining the optimal matching point was difficult. Thus, a hierarchical searching method was used here. While the searched objects had different significance to matching points, the weight was given by distance to determine the final matching height. Finally, real cloud heights were determined according to the matching method of cloud shadow similarity. We used five images of HJ-1B CCD and IRS data in the Heihe area on June 8, 2012. From all the matched results, the corresponding regions of the matched results had high cloud heights where thermal infrared data also had high cloud heights. At the same time, resolution and details were improved. To evaluate the accuracy of the calculated heights, we selected 10 highly recognizable feature shadow points in each image and marked their coordinates in the image. We set the actual shadow points as reference and calculated the offset of the same feature shadow point, in which we could obtain the error of each cloud height. We found that the errors between 0.1 km to 0.3 km were 70% among all 50 points, 12% were less than 0.1 km, and 0.25 km was the average error of all points. Compared with other studies, cloud height accuracy was higher in our study. Also, we chose SIFT algorithm to match cloud heights by using two types of data and compared the matched results of some feature points with our method. The height accuracy obtained by SIFT was lower than that of our algorithm. In addition, unlike some feature point matching methods, our method can complete full-information matching. A cloud edge height matching method based on Euclidean distance transform by hierarchical searching is proposed in our paper. The method can match the cloud height information of low-resolution thermal infrared to corresponding cloud edge of high-resolution image. Experimental results showed that the matched results followed the distribution and variation law of thermal infrared cloud heights, as well as the cloud edge heights with high accuracy and detailed characteristics. To a certain extent, our study solved the resolution problem in obtaining cloud height by thermal infrared data. In addition, compared with some matching methods of feature points, our method could complete full-information matching and had a higher accuracy. However, the accuracy of matched results would be influenced by many factors, such as accuracy of cloud detection, surface in homogeneity, and image registration. The method in our paper is only for cloud edge height. Thus, matching for other parts of cloud still needs further research.

# Key words

cloud edge heights , Euclidean distance transform (EDT) , contour scanning , cloud shadows

# 2 基于欧氏距离的围线搜索云边缘高度匹配方法

## 2.2 针对背景图像的距离变换

 $d_p \left[({{x}_{p}}\text{，}\!\!{{y}_{p}})\text{，}({{x}_{i}}\text{，}\!\!{{y}_{i}})\right]={{\left[ {{\left({{x}_{p}}-{{x}_{i}} \right)}^{2}}+{{\left({{y}_{p}}-{{y}_{i}} \right)}^{2}} \right]}^{{}^{1}\!\!\diagup\!\!{}_{2}\;}}$ (1)

## 2.3 围线搜索与距离加权确定匹配点云高

(1) 初始围线半径 r设定为1.5个像素宽，将围线区内的像素集合作为候选点集 C =[ c k ]( k=1， $\cdot \! \cdot \! \cdot$q)，若集合内候选点数量 q大于2且小于8时，则将每个候选点距目标点的距离 d k 求和得 Sd p ，并求每个候选点相对目标点的距离权值 w k ，最后将高度 h k 加权计算并赋值给目标点，得到距离加权后的高度值 H p

 $S{d_p} = \sum\limits_{k = 1}^q {{d_k}}$ (2)
 ${{w}_{k}}={{{d}_{k}}}/{S{{d}_{p}}}\;$ (3)
 ${H_p} = \sum\limits_{k = 1}^q {{w_k} \times {h_k}}$ (4)

(2) 若集合内候选点数量 q小于2或大于8时，则跳过，继续扩大 r。至7.5个像素宽之前，以0.5为步长由内向外逐步扩大围线区范围，8到12以1个像素宽为步长，直至满足条件。重复步骤(1)中的计算过程。

(3) 逐一选取目标像素集 P 中的点完成上述步骤。若某点的 r>12仍未满足条件，则终止扫描，跳至下一个目标点。

## 2.4 迭代法确定真实云高度

 \begin{align} & {{H}_{\min }}=\max \left(0.2\text{，}{{{T}_{surface}}-{{T}_{cloud}}}/{9.8}\; \right)\\ & {{H}_{\max }}=\min \left(12\text{，}{{{T}_{surface}}-{{T}_{cloud}}}/{1}\; \right)\\ \end{align} (5)

# 3 结果及分析

## 3.2 云高的准确性评价

Table 1 Error of cloud heights of feature cloud shadow points

 特征点编号 图像1 图像2 计算阴影坐标 实际阴影坐标 高度误差/m 计算阴影坐标 实际阴影坐标 高度误差/m 1 (254，91) (251，89) 217 (145，90) (142，94) 300 2 (260,101) (257,101) 180 (145，85) (142，89) 300 3 (226,236) (224234) 170 (162，64) (161，63) 85 4 (241,225) (238,221) 300 (167，64) (167，63) 60 5 (191,340) (188,339) 190 (188，65) (186，62) 217 6 (115,256) (112,254) 217 (215，63) (214，62) 85 7 (115,267) (112,270) 256 (305,281) (303,283) 170 8 (42,334) (37,333) 307 (310,280) (311,280) 60 9 (376，48) (371，43) 427 (47,114) (45,116) 170 10 -306215 -302214 249 (170,255) (169,254) 85

Table 2 Comparison for matched cloud heights

 /km 算法 特征点云高 图像1 图像2 特征点 1 2 3 4 5 1 2 3 4 5 SIFT 1.7 1.8 2.3 1.6 2.3 2.6 2.1 2.1 2.2 1.9 本文算法 2.6 2.6 2.9 2.5 2.8 4.1 4 3.9 3.8 4.3